Skip to main content
PLOS ONE logoLink to PLOS ONE
. 2024 Jul 31;19(7):e0306549. doi: 10.1371/journal.pone.0306549

A deep learning framework for predicting endometrial cancer from cytopathologic images with different staining styles

Ruijie Wang 1, Qing Li 2, Guizhi Shi 3, Qiling Li 2, Dexing Zhong 1,4,5,*
Editor: Kazunori Nagasaka6
PMCID: PMC11290691  PMID: 39083516

Abstract

Endometrial cancer screening is crucial for clinical treatment. Currently, cytopathologists analyze cytopathology images is considered a popular screening method, but manual diagnosis is time-consuming and laborious. Deep learning can provide objective guidance efficiency. But endometrial cytopathology images often come from different medical centers with different staining styles. It decreases the generalization ability of deep learning models in cytopathology images analysis, leading to poor performance. This study presents a robust automated screening framework for endometrial cancer that can be applied to cytopathology images with different staining styles, and provide an objective diagnostic reference for cytopathologists, thus contributing to clinical treatment. We collected and built the XJTU-EC dataset, the first cytopathology dataset that includes segmentation and classification labels. And we propose an efficient two-stage framework for adapting different staining style images, and screening endometrial cancer at the cellular level. Specifically, in the first stage, a novel CM-UNet is utilized to segment cell clumps, with a channel attention (CA) module and a multi-level semantic supervision (MSS) module. It can ignore staining variance and focus on extracting semantic information for segmentation. In the second stage, we propose a robust and effective classification algorithm based on contrastive learning, ECRNet. By momentum-based updating and adding labeled memory banks, it can reduce most of the false negative results. On the XJTU-EC dataset, CM-UNet achieves an excellent segmentation performance, and ECRNet obtains an accuracy of 98.50%, a precision of 99.32% and a sensitivity of 97.67% on the test set, which outperforms other competitive classical models. Our method robustly predicts endometrial cancer on cytopathologic images with different staining styles, which will further advance research in endometrial cancer screening and provide early diagnosis for patients. The code will be available on GitHub.

Introduction

Endometrial cancer is one of the most common tumors in the female reproductive system and usually occurs in postmenopausal women [1, 2]. It is the leading cause of cancer-related deaths in women worldwide [3], with approximately 76,000 deaths each year [1]. And the incidence and mortality of endometrial cancer is expected to continue to rise in the coming decades [4, 5]. Studies have shown that endometrial screening can help to detect cellular lesions, and improve long-term patient outcomes [6, 7]. It would significantly improve survival rates [8]. So, endometrial cancer screening is crucial.

However, there are few tools available for the endometrial cancer screening. A minimally invasive method based on cytopathology to address endometrial cancer screening is a hot topic of current research and future development [9]. And it has been widely used in countries such as Japan [10, 11]. Moreover, it is considered to be cost-effective and more useful for early screening than invasive endometrial biopsy and hysteroscopy [1215]. Nevertheless, there are still many difficulties in advancing cytopathological screening.

Firstly, there are no endometrial cytopathology datasets that contain segmentation and classification labels, due to the difficulty of data acquisition and the high cost of high-quality annotation. To combat the challenge, our team collecting 139 cytopathology whole slide images (WSI) with our own designed endometrial sampling device Li Brush (20152660054, Xi’an Meijiajia Medical Technology Co., Ltd., China). Among them, 39 WSIs are papanicolaou stained, and 100 WSIs are hematoxylin and eosin (H&E) stained. These WSIs are annotated by two cytopathologists, thus building a dataset for cytological screening of endometrial cancer. To the best of our knowledge, this is the first cytopathology dataset that includes segmentation and classification labels.

Secondly, diagnosing cytopathological slides is a time-consuming and complex task [16]. Subjective discrepancies and heavy workloads affect the productivity of cytopathologists [17]. As a powerful tool, deep learning can provide objective references for doctors and further improve their work efficiency [18, 19]. Therefore, it is widely used in thyroid cancer [20], cervical cancer [21], liver cancer [22, 23], and other diseases to improve the diagnostic efficiency [24]. And in endometrial diagnosis, deep learning is usually used for segmentation and classification tasks. In the field of segmentation, Erlend Hodneland et al. used a UNet-based 3D convolutional neural network (CNN) to segment endometrial tumors on radiology images, which aimed at generating tumor models for better individualized treatment strategies [25]. Zhiyong Xia et al. designed a dense pyramidal attention U-Net for hysteroscopic images and ultrasound images, which can help doctors to accurately localize the lesion site [26]. In addition, segmentation algorithms are often used to assist in confirming the depth of myometrial infiltration in endometrial cancer [2730].In the field of classification, Christina Fell et al. used CNNs to classify endometrial histopathology images at the WSI level, which are categorized as “malignant”, “other or benign” and “insufficient” [31]. Sarah Fremond et al. proposed an interpretable endometrial cancer classification system, which can further predict four molecular subtypes of endometrial cancer through self-supervised learning [32]. Min Feng et al. develop a deep learning model for predicting lymph node metastasis from histopathologic images of endometrial cancer, which is believed to predict metastatic status and improve accuracy [33]. In summary, In summary, we note that deep learning models are commonly used for radiology images [34, 35] and histopathology images [3638] in endometrial studies. Therefore, there is still a lack of endometrial cancer screening algorithms based on cytopathologic images. The aim of our study is to develop a new algorithm that learns cytopathological features and provides an assisted diagnostic strategy.

Here we propose an innovative two-stage framework for endometrial cancer screening. In this study, we found that the staining styles of slides was performed differently in different medical centers [39]. Some endometrial samples were stained with H&E, while others were stained with papanicolaou. In addition, the stained slides can also be highly variable due to the preservation environment, changes in the scanner, etc [39]. This can affect the final diagnosis results [40, 41]. Therefore, we have improved the automated screening framework to increase its robustness and accuracy.

In clinical diagnosis, cell clumps are regions of interest (ROIs) for cytopathologists, while the background contains unnecessary noises [42]. So, in the first stage, we propose an improved segmentation network CM-UNet, which extracts ROIs from cytopathology images. We introduce a channel attention (CA) module and a multilevel semantic supervision (MSS) module to obtain more local and global contextual representations. In addition, we added novel skip connections to efficiently extract multi-scale features.

In the second stage, we need to classify ROIs to screen positive cell clumps. Since the obtained ROIs vary in shape and size, the different representations among these ROIs may affect the performance of the classification model. We propose the contrastive learning based algorithm ECRNet to classify ROIs. In contrast to current contrastive learning methods that treat different augmentations of the same image as positive pairs, we introduce the label memory bank to preserve the representation information of the image and the corresponding labels. ECRNet treats two instances with the same label as a positive sample pair, and two instances with different labels as a negative sample pair. Thereby, different images with the same semantics are better aggregated in the representation space, while negative sample pairs are separated in the representation space. This makes better use of class-level discriminative information and enhances the generality of the algorithm to some extent.

Finally, our experimental results show that the two-stage framework performs well on cytopathology image with different staining styles. The framework can accurately present negative and positive cell clumps to cytopathologists, providing objective decision support.

The main contributions of our study are as follows:

  1. Computer-aided diagnostic studies for endometrial cancer screening are scarce and there is a lack of available datasets. Therefore, our team created an endometrial cancer cytology dataset, which was annotated by two cytopathologists. This dataset contains segmentation labels and classification labels that can be used for deep learning.

  2. Compared to histopathology image segmentation, endometrial cytology images have more noise and sparser semantic features, which pose a challenge to segmentation algorithms. We propose a segmentation model based on the UNet architecture, and for better extraction of semantic features in cytology images, we introduce the CA module and the MSS module to learn more local and global contextual representations.

  3. Considering that images with the same classification label may be represented differently from each other, e.g., variations in staining styles, which may affect the classification model performance. In order to make full use of the image content information and label information, we propose ECRNet and introduce the label memory bank to make ECRNet focus more on the class-level discriminative information.

  4. The framework performs efficiently on H&E-stained and papanicolaou-stained cytopathology images, and shows cytopathologists the negative and positive cell clumps. On the test set, it achieved an average accuracy of 98.50%, an area under the curve (AUC) of 93.66% compared to other classical models. The results show that our model can contribute to medical decision-making.

The rest of this paper is organized as follows: Section 2 describes the materials and methods; Section 3 analyzes our results; Section 4 and Section 5 discuss and conclude our work, respectively.

Methods

Data collection

From July 2015, we collected endometrial cells using a sampling device of our own design, the Li Brush (20152660054, Xi’an Meijia Medical Technology Co., Ltd., Xi’an, China). This hospital routine work lasted for seven years since 2015 (XJTU1AHCR2014-007). Until after 2019, endometrial cells obtained from the Li Brush were used in this study (XJTU1AFCRC2019SJ-002). It is important to note that our team spent seven years collecting endometrial cells. However, all data that was used for analysis was obtained after 2019. Therefore, no retrospective ethical approval was involved. The endometrial cells collected from 03/12/2019 to 03/12/2020 used in this study were done so under IRB approval.

From 2019 to 2020, our team collecting images. After 2020 and up to 2022, we are mainly working on the annotation process, and building the endometrial cytopathology image dataset. 139 women who underwent curettage or hysterectomy at the First Affiliated Hospital of Xi’an Jiaotong University were registered in the Obstetrics and Gynecology Registry. Patient exclusion criteria were as follows: (1) diagnosed with suspected pregnancy or pregnancy; (2) diagnosed with acute inflammation of the reproductive system; (3) patients who had undergone hysterectomy for a previous diagnosis of cervical cancer, cervical intraepithelial neoplasia, or ovarian cancer and so on; (4) diagnosed with dysfunctional clotting diseases; and (5) women who body temperature at 37.5°C or higher twice a day were also excluded.

The study is approved by the Ethics Committee of the First Affiliated Hospital of Xi’an Jiaotong University (XJTU1AFCRC2019SJ-002), and written consent was obtained from all patients. Minors were not included in the study. And the authors will not have access to information that could identify individual participants. After three years of collection, 139 patients are eventually included in the study and their details are shown in Table 1, which includes the age of the patients, childbirth history, menstrual status, and any other diseases. The detailed information of age distribution is shown in Fig 1. In addition, histopathological diagnosis was also collected and the specific information is shown in Table 2. It is worth mentioning that the protocols used in this study were all in accordance with the ethical principles of the Declaration of Helsinki on Medical Research [43].

Table 1. Patient characteristics.

Characteristics Number
SOURCE
 IPD 81
 OP 58
AGE
 <40 years old 32
 ≥40 years old 107
MENSTRUAL STATUS
 Premenopausal 77
 Postmenopausal 35
 AUB 27
OTHER DISEASE
 Ovarian cancer 18
 Hypertension 24
 Diabetes 24
 Hormone replacement therapy 32
CHILDBIRTH EXPERIENCE
 Yes 101
 No 28

IPD, Inpatient Department. OP, Outpatient. AUB, Abnormal uterus bleeding. Some information of the patients is missing.

Fig 1. The age distribution of the patients.

Fig 1

Table 2. Pathological diagnosis.

Histological diagnostic results Number
Proliferative endometrium 14
Secretory endometrium 8
Atrophic endometrium 10
Mixed endometrium 2
Endometrial hyperplasia without atypia 39
Endometrial atypical hyperplasia 4
ENDOMETRIAL CARCINOMA 62
 Endometrioid carcinoma, G1/G2 47
 Endometrioid carcinoma, G3 11
 Serous carcinoma 2
 Clear cell carcinoma 2

G1, G2, G3 represent grade 1, grade 2, grade 3 respectively.

Our study was based on all the cases collected from 2019 to 2020. The data in this work was cleaned so that it did not contain private patient information. The datasets used or analysed during the current study are available from the corresponding author on reasonable request.

We collected endometrial cells with Li Brush, and H&E staining or papanicolaou staining was used for liquid-based cytology specimens of endometrial cells. Histopathological diagnosis of the same patient was also collected at the same time. When the cytopathologic diagnosis was consistent with the histopathologic diagnosis, the case was included in the study. Finally, 39 whole slide images (WSIs) are papanicolaou stained and 100 WSIs are H&E stained.

The MOTIC digital biopsy scanner (EasyScan 60, 20192220065, China) with an ×20 lens was used to scan cytopathological slides in the counterclockwise spiral. And the focal length is automatically adjusted. Since each WSI is very large (e.g., 95200 × 87000 pixels), which cannot be directly input to the deep learning model. We crop the WSIs into same-sized images (1024 × 1024 pixels). Then a simple but effective thresholding algorithm is used to remove the meaningless background images. Specifically, we first calculate the mean and standard deviation of each image in RGB space. And then, those images with mean between 50 and 230 and standard deviation above 20 are retained. These images often contain meaningful cell clumps.

And the image annotation process includes segmentation label annotation and classification label annotation. Segmentation labels were obtained by two experienced pathologists using Adobe Photoshop CC (2019 v20.0.2.30). First, one senior cytopathologist segmented the cell clumps and the results were reviewed by the other cytopathologist. After the review is accurate, the pathologists annotated the cell clumps according to the International Society of Gynecologic Pathologists and the 2014 World Health Organization classification of uterine tumors. All cell clumps were classified into two categories: malignant (atypical cells of undetermined significance, suspected malignant tumor cells, and malignant tumor cells), and benign (non-malignant tumor cells). Benign diagnosis is defined as cell clumps with neat edges, nuclei with oval or spindle shape, and evenly distributed, finely granular chromatin. Malignant diagnosis referred to a three-dimensional appearance, irregular (including dilated, branched, protruding, and papillotubular) edge, with the nucleus poloidal disordering or disappearing (including megakaryocyte appearance, nuclear membrane thickness, and coarse granular or coarse block chromatin). Both benign and malignant tumors were followed up histologically. Undoubtedly, the cell clumps in the negative slides are all negative, but the ones in the positive slides have both negative and positive cell clumps. Therefore, the two cytopathologists again vote on the labeling of each cell clump, when the votes do not agree, they will discuss it. If the discussion fails to result in an accurate diagnosis, the cell clump is discarded. These measures ensure the accuracy and consistency of segmentation and classification labels.

Based on the results annotated by pathologists, we established the XJTU-EC dataset, which containing 3,620 positive images (endometrial cancer cell clumps and endometrial atypical hyperplasia cell clumps) and 2,380 negative images.

Endometrial cancer screening

In this paper, we propose a novel framework for endometrial cytology image analysis, applying two fully convolutional networks for early diagnosis. Firstly, a dense connection-based semantic segmentation network, with CA and MSS modules, is used for extract ROIs; secondly, a model based on contrastive learning is applied to classify ROIs. The final results confirm the effectiveness of this strategy. All details as shown in Fig 2.

Fig 2. The proposed pipeline for cancer screening using endometrial cytology data.

Fig 2

CM-UNet

To eliminate the interference of neutrophils, dead cells and other impurities contained in the background, and helping the pathologist to better localize the lesion, the first step is to segment the endometrial cell clumps.

The CM-UNet performed ResNet101 as the backbone, and computes the aggregation of all feature maps at each node by applying dense connections [4446]. As shown in Fig 3.

Fig 3. Pipeline of the proposed framework for cell clumps segmentation.

Fig 3

CM-UNet allows more flexible feature fusion at decoder nodes through densely connected skip connections. L is the loss function. The bold links represent the necessary depth supervision and the light coloured links represent the optional ones.

Let xi,j denote the output of node XI,J, then xi,j can be represented as follows:

xi,j=H(D(xi1,j)),j=0Hxi,kk=0j1,U(xi+1,j1),j>0 (1)

where H(·) is a convolution operation followed by an activation function, D(·) and U(·) denote a down-sampling layer and an up-sampling layer respectively, and [·] denotes the concatenation layer.

To better accommodate different image staining styles, inspired by [47], we apply CA module at the bottleneck of the encoder-decoder network. The CA module integrates the semantic relationships between different channel mappings, and emphasises strongly interdependent channel mappings by adjusting the weights [48]. As shown in Fig 4.

Fig 4. The details of the channel attention module.

Fig 4

We perform matrix multiplication between the feature maps x4,0 and the transpose of x4,0. Then apply a softmax layer to calculate the influence of the ath channel on the bth channel:

Mba=exp(xa4,0xb4,0)a=1Kexp(xb4,0xb4,0) (2)

where K is the number of channels. On this basis, the final output E is described as follows:

Eb=βa=1K(Mbaxa4,0)+xb4,0 (3)

the scaling parameter β is gradually learns a weight from 0, which is updated in subsequent learning.

To solve the gradient disappearance/explosion problem, we introduce the MSS module [49]. By setting appropriate weights for different side output layers, a deeper semantic representation is learned. Assuming that the original input image is represented as R, and d represents the depth of the CM-UNet model, each output layer S performs a 1 × 1 convolution operation, followed by the global average pooling to extract global contextual information. And then we assign a weight factor α for each layer of the model. So, the side loss can be defined as:

Lside=i=1dαiSi(R,θi) (4)

where θi denotes the relevant parameter of the ith output layer. And α3, α2, α1, α0 is set sequentially to 0.1, 0.3, 0.6, 0.9.

In addition, we introduce a hybrid segmentation loss LT,P to address the class imbalance in the segmentation task:

LT,P=1Nc=1Cn=1N(tn,clogpn,c+2tn,cpn,ct2n,c+p2n,c) (5)

where tn,cT, and pn,cP denote the ground truth and predicted label for class c and nth pixel in the batch. T represents the ground truth, and P represents the prediction probability. C represents the number of categories, and N represents the number of pixels in one batch.

Ultimately, the overall loss function of CM-UNet is defined as the weighted sum of the hybrid segmentation loss LT,P and the side loss Lside. The final loss function is shown as below:

L=i=1d(LT,P+Lside) (6)

where d represents the depth of the CM-UNet model.

We trained the segmentation network using the dataset annotated with pathologists, and performed ten-fold cross-validation. It should be noted that segmentation results often have some holes and flaws. Gaps and noise in the images are eliminated by morphological processing. Finally, we obtained a ROI dataset consisting of cell clumps for the next stage.

ECRNet

After segmentation of the cytopathological images, noise such as single cells and leukocytes are removed from the background, leaving only the ROIs, as shown in Fig 2. Next, we filled its surroundings with pixels of value 0 until the size was adjusted to 512 × 512, in order to further the image analysis task.

Due to the complexity of the endometrial cell features, there is an urgent need for a powerful deep learning classifier to learn and classify cell features. We propose a state-of-the-art method based on contrastive learning to address the above needs, and named ECRNet. The details are shown in Fig 5.

Fig 5. The details of ECRNet.

Fig 5

ECRNet consists of two parts: contrastive learning and supervised learning. In contrastive learning, we want to import classification labels in the training data to improve the performance of the classification task. Therefore, we introduce the label memory bank [50]. Two instances with the same label are considered as the same pair, while two instances with different labels are considered as different pairs. This process can be considered as a dictionary look-up task. Given an encoding query q (with label y), we need to look up the corresponding positive key k from a dictionary. Assume that a dictionary of n encoded labeled keys {(k1, y1), (k2, y2),…, (kn, yn)}, for the given encoded query (q, y), its label contrastive loss Lcon can be calculated as:

Lcon=logi=1nIIyi=yexp(sim(q,ki)/τ)i=1nexp(sim(q,ki)/τ) (7)

where II is an indicator function that takes the value of 1 if y (the label of the q) and yi (the label of the ki) are the same, otherwise it is 0. sim(·) is a similarity function, and τ is a temperature parameter.

In order to store the large number of image representations and labels in the label memory bank, we introduce the momentum update method so as to dynamically construct a large and consistent dictionary [51]. This not only reduces the computational overhead, but also allows the learned representations transfer well to the downstream task. The specific update equation is as follows:

θkmθk+(1m)θq (8)

Here m∈is a momentum coefficient, taking values between [0,1). θk is the parameter of the encoder fk, θq is the parameter of the encoder fq.

In supervised learning, we choose the VGG-16 as the classifier [52, 53]. The cross-entropy loss function is our classifier loss functions and defined as follows:

Lcla=1QqiQjYIIqi,jlog(pqi,j) (9)

where Q is the set of the query image representations, Y is the label set, pqi,j is the predicted probability that the query image qi is predicted to be j. II is an indicator function that takes the value of 1 if the query image qi is classified correctly, otherwise it is 0.

Finally, the ECRNet total loss is calculated as follows:

Ltotal=Lcla+βLcon (10)

where β is a hyperparameter to adjust the relative weight between classification loss and contrastive loss. The value of β in general is 0.5.

Evaluation metrics

We chose the Dice coefficient to evaluate the segmentation model. It is a measure of the similarity between two samples and is one of the commonly used evaluation criteria for segmentation [54]. When the Dice coefficient is 1, it means that the segmentation model achieves perfect results. The Dice coefficient is then calculated as follows:

Dice=2TPT+P (11)

where T is the set of ground truth, and P is the set of corresponding segmentation results, respectively.

To better evaluate the performance of the classification model, we use four commonly used quantitative indicators of accuracy, sensitivity, specificity, and F1-Score as the evaluation indicators of the classification model. The indicators are defined as follows:

Accuracy=TP+TNTP+FP+TN+FN (12)
Sensitivity=TPTP+FN (13)
Precision=TPTP+FP (14)
F1_Score=2Precision×SensitivityPrecision+Sensitivity (15)

where TP, TN, FP and FN represent true positives (correctly classified as positive), true negatives (correctly classified as negative), false positives (incorrectly classified as positive) and false negatives (incorrectly classified as negative), respectively.

In addition, in order to compare the different performance of various classifiers, we select ROC curve and AUC value to visualize the classification results of each classifier. The ROC curve graph reflects the relationship between sensitivity and specificity. Its abscissa represents FPR (false positive rate), and the ordinate is called TPR (true positive rate). The AUC value can be obtained by calculating the area under the ROC curve. A higher AUC value can prove the superiority of the classification model.

Experiments and results

Implementation details

The cytopathology images are all augmented by vertical flipping, horizontal flipping, random rotation (90°, 180°, 270°), scaling and graying to improve the framework performance. We use the ImageNet 25 pre-trained weights as the encoder weights to initialize the segmentation and classification models, respectively, while the weights for the decoder part are randomly initialized. And the Adam optimizer is introduced to optimise the model, with an initial learning rate of 5 × 10−3 [55]. The temperature parameter τ is 0.07, the momentum parameter m is 0.9. Finally, the training batch size is set to 32.

We used ten-fold cross-validation to test our models. All networks are implemented based on the TensorFlow framework, and trained by two GPU cards (NVIDIA GeForce GTX 1080), with Python 3.6.12(Python Software Foundation, Wilmington, DE, USA), keras 2.4.3 (Google Brain, Mountain View, CA, USA) and TensorFlow 2.2.0 (Google Brain, Mountain View, CA, USA).

Segmentation results

We applied our segmentation algorithm and other classical segmentation algorithms, such as fully convolutional networks (FCN) [56], UNet [44], UNet++ [45], LinkNet [57], DeepLabV3 [58], and DeepLabV3+ [59] on HE-stained images and papanicolaou-stained images, respectively. The experimental results are shown in Table 3. Our model shows great potential in segmenting cell clumps, with the average Dice value exceeding 0.85. In addition, we conducted ablation experiments, as shown in Table 3, to verify the role played by the CA module and the MSS module in the segmentation model.

Table 3. Comparison experiments.

Model Dice Training time Inference time Params
FCN 0.61 2.17h 0.045s 270M
UNet 0.75 2.50h 0.022s 33M
UNet++ 0.85 1.49h 0.029s 30M
LinkNet 0.79 1.21h 0.030s 12M
DeepLabV3 0.81 2.30h 0.030s 54M
DeepLabV3+ 0.85 1.80h 0.025s 41M
UNet++
(with CA)
0.86 1.52h 0.030s 31M
UNet++
(with MSS)
0.88 1.88h 0.039s 33M
CM-UNet 0.89 1.90h 0.039s 33M

Comparison experiment of our segmentation algorithm with other classical segmentation algorithms. The inference time is calculated by the single image.

The segmentation results are shown in Fig 6. The first column is the cytological image, the second column is the result of FCN, the third column is the result of UNet, the fourth column is the result of UNet++, the fifth column is the result of LinkNet, the sixth column is the result of DeepLabV3, the seventh column is the result of DeepLabV3+, the eighth column is the result of CM-UNet, and the ninth column is the ground truth annotated by the pathologist. The red boxes represent the over-segmented area, and the yellow boxes represent the under-segmented area. As can be seen from this figure, FCN and UNet have more under-segmentation and fail to identify all cell clumps, which is not suitable for cytopathology image segmentation. Whereas LinkNet and DeepLabV3 tend to over-segment, mistaking mucus and single cells for cell clumps, which does not benefit the subsequent classification task and is therefore also not applicable. The segmentation results of UNet++, DeepLabV3+ and CM-UNet basically conform to the gold standard. However, UNet++ performed moderately well on H&E-stained images but poorly on papanicolaou-stained images, occasionally missing cell clumps. DeepLabV3+, on the other hand, made fewer errors on the papanicolaou-stained images but missed cell clumps on the H&E-stained images. Taken together, the segmentation result of CM-UNet is closer to the annotation of pathologists. It is able to segment all cell clumps and extract ROIs. This demonstrates the performance of our segmentation network.

Fig 6. Comparison with classical segmentation algorithms.

Fig 6

We randomly show the segmentation results of four H&E-stained and four papanicolaou-stained cytopathology images. The red boxes represent the over-segmented area, and the yellow boxes represent the under-segmented area.

Classification results

We input the ROIs into ECRNet for ten-fold cross-validation. Table 4 shows the results of the ablation experiments. The backbone (VGG-16) classified the extracted ROI dataset with an accuracy of 91.07%. In contrast, VGG-16 with contrastive learning component achieves 7.43% higher accuracy than backbone.

Table 4. Ablation experiments.

Model Accuracy (%) Precision (%) Recall (%) F1-score (%)
VGG-16
(One stage)
84.29 83.23 88.74 85.90
VGG-16
(Two stage)
91.07 90.38 93.38 91.86
ECRNet
(One stage)
89.17 88.03 90.67 89.33
ECRNet
(Two stage)
98.50 99.32 97.67 99.33

Table 4 also shows the importance of the segmentation strategy. In the two-stage strategy, we first segment the cytopathology images to obtain ROIs, and apply the classifier to classify the ROIs. In the one-stage strategy, the classifier directly classifies the cytopathology images containing the background. All data sizes are scaled to 512 × 512. The results of the experiment. Among them, VGG-16 performed the worst under the one-stage strategy, with an accuracy of 84.29%. In contrast, ECRNet performed best under the two-stage strategy.

As shown in Table 5, we compared ECRNet with five classical deep learning models. There are the MobileNet [60], the ResNet-101 [61], the Inception-V3 [62], the ViT [63], the ResNeXt-101 [64], the EfficientNet [65], the DenseNet-121 [66], and VGG-16 [67]. Note that all network parameters remain the same as previously described, and the initialization weights are the ImageNet pre-trained weights. These results are obtained in the two-stage framework, which is based on the classification of cell clumps.

Table 5. Comparison with baseline methods.

Model Accuracy (%) Precision (%) Recall (%) F1-score (%) Params
VGG-16 +SVM 78.83 75.62 78.15 76.86 138M
ResNet-101 +SVM 88.55 84.30 85.71 85.00 24M
Inception-V3 +SVM 87.60 80.77 88.24 84.34 22M
MobileNet-V1 82.99 79.11 74.79 76.89 5M
Inception-V3 82.17 81.43 83.33 82.37 22M
ViT 65.00 62.43 95.58 75.52 343M
ResNeXt-101 86.50 81.16 99.12 89.24 79M
EfficientNet-B7 88.50 88.79 91.15 89.96 66M
ResNet-101 92.17 97.03 87.00 91.74 24M
DenseNet-121 93.50 92.23 95.00 93.59 8M
Ours 98.50 99.32 97.67 99.33 138M

In this part, we found that the ViT model performs poorly, which may be due to the small size of our dataset and overfitting of the ViT model. In addition, the MobileNet-V1 model also performs poorly, which may be due to the fact that lightweight networks are not good at learning complex cytopathological features. In contrast, the Inception-V3, the ResNeXt-101, the EfficientNet-B7, the ResNet-101, the DenseNet-121, and the ECRNet models performed better, with mean values of accuracy, precision, recall, and F1-scores above 80%. Specifically, ResNeXt-101 has the highest recall, but has 12% less classification accuracy than ECRNet. And ResNet-101 has a precision of 97.03%, second only to ECRNet, which indicates that it has a lower false positive rate. However, the recall of ResNet-101 is only 87.00%, which indicates that it has a higher false negative rate. And it is more likely to miss cancer cell clumps. And ECRNet has the best classification performance, outperforming other classification models in terms of accuracy, precision and F1-score.

Furthermore, several studies [68] show that using CNNs to extract features and train linear support vector machines (SVMs) achieves better performance than end-to-end CNN-based classifiers. Therefore, we use three classical CNNs, that is VGG-16, ResNet-101 and Inception-V3, to obtain the image feature vectors. These feature vectors are then used to train a linear SVM that classifies all ROIs as positive or negative. As the dimensionality of the feature maps is large, we used principal component analysis to reduce the dimensionality of the image features. The SVM classifier uses a radial basis function kernel with parameters γ and C set to 0.0078 and 2, respectively. And the rest of the experimental settings are consistent with those described previously. The results are shown in Table 5. As can be seen, all results are unsatisfactory, with VGG-16+SVM having an accuracy of only 78.83%.

We further plotted the ROC curves for the classifiers in the binary classification task. As can be seen from Fig 7, ECRNet outperformed all the models. This experiment suggests that the model benefits more from two-stage framework than one-stage framework. Our two stage strategy is effective.

Fig 7. The false positive rate, the y-axis represents the true positive rate.

Fig 7

The points above the diagonal line are indicated by dashed lines indicating a better than random classification result, i.e. an AUC value of 0.5.

Finally, we also discuss the ECRNet classification failure cases. As shown in Fig 8, we randomly listed 8 correctly classified cell clumps and 8 failure cases in the test set. Of these, the 4 false-negative (missed diagnosis) cases consisted of 1 well-differentiated endometrial adenocarcinoma and 3 poorly differentiated endometrial adenocarcinomas. In contrast, 4 false-positive (over diagnosed) cases included 3 normal cell clumps that were classified as cancer. In some of these classification failure samples, it was difficult for the classification model to extract deeper features because of cell stacking and obscure structural features. In addition, another part of the failed cases showed that the number of cells in the image was small, which was easier to be misclassified by the classification model. This suggests that ECRNet’s ability to classify small targets needs to be strengthened in future work.

Fig 8. Examples of correct and incorrect predictions of ECRNet.

Fig 8

In addition, we conducted an external validation of the paper’s algorithm using the public dataset. The externally validated data were obtained from the public data platform, AIstudio, accessible via the Internet [69]. It is worth noting that the data used for external validation came from the public dataset and did not contain segmentation labels, so this external validation mainly evaluated the classification performance of ECRNet.

The dataset used for external validation consisted of 848 negative endometrial cytopathology images and 785 positive endometrial cytopathology images. The ratio of negative to positive images was 1.08:1. All images are papanicolaou stained images. During this external validation, we also chose four classification models (the ResNet-101 model, the DenseNet-121 model, the EfficientNet-B7 model, and the ResNeXt-101 model) to compare with our model. This is because these four models perform well in the classification task, second only to ECRNet. Specifically, these models above were first trained using images (1024 × 1024 pixels) from the XJTU-EC dataset instead of cell clumps (ROIs), and then tested directly on an external dataset. The specific results are shown in Table 6 as following:

Table 6. External validation comparison results.

Model Accuracy (%) Precision (%) Recall (%) F1-score (%)
ResNeXt-101 64.50 86.20 44.20 58.44
EfficientNet-B7 83.53 78.72 59.2 67.58
ResNet-101 73.50 100.00 53.10 69.37
DenseNet-121 80.10 77.20 67.20 71.90
Ours 95.32 94.57 96.17 95.37

In this external validation experiment, ResNet-101 has the highest precision of 100%, which means it has no false positives in the external validation. However, both ResNeXt-101, EfficientNet-B7 ResNet-101, and DenseNet-121 have low recall, which means they can easily miss screening positive patients. This is unacceptable for clinical tasks. In contrast, our model achieved the highest recall of 96.17%. In addition, ECRNet has the highest accuracy of 95.32%, followed by the DenseNet classifier with 83.53%. It is worth noting that ECRNet significantly outperforms the other four classifiers in terms of the F1 score. The F1 score is the reconciled average of precision and recall, which combines the information of precision and recall to provide a more comprehensive assessment of the performance of the classifiers, and the higher the F1 score, the better the performance of the classifiers. The higher the F1 score, the better the performance of the classifier. In summary, ECRNet has the best performance in this external validation experiment.

Discussion

Currently, there is no well-established method to screen endometrial cancer. The main screening tests for endometrial cancer include ultrasound, hysteroscopy and endometrial biopsy. Sequential transvaginal ultrasound scan is a less invasive method of assessment, but lacks a high degree of specificity. Until now, the collection of tissue samples from the endometrium and analysis of histopathological images by physicians has been the gold standard for the diagnosis of endometrial cancer. However, both endometrial biopsy and hysteroscopy are invasive and require the cooperation of anaesthetists, which is expensive. As a result, cytopathology-based screening for endometrial cancer is becoming increasingly desirable.

Due to the lack of relevant data and the complexity of cell morphology, endometrial cancer screening based on cytopathology is difficult to promote. Therefore, our team spent three years collecting and annotating WSIs from 139 patients to create the endometrial cytopathology image dataset, named the XJTU-EC dataset. Since our dataset contains both papanicolaou and H&E stained images, which are the most common staining modalities for cytology images. Therefore, it can be somewhat considered a representative dataset. In addition, the data includes patients of different age, so the XJTU-EC dataset has more diversity. Based on this dataset, we investigated the first clinically automated deep learning framework for extracting and identifying normal or cancerous endometrial cell clumps. The results will be presented to cytopathologists as a reference.

In order to adapt to different staining styles, in the cell clump extraction stage, we use the robust UNet as the backbone, which has been previously generalized to many datasets. Based on this, we introduced the CA module pay attention to global contextual information, and MSS module to aggregate semantic features at multiple scales. So, our method achieves better segmentation results. Experiments demonstrate that CM-UNet is able to perform well on both H&E-stained images and papanicolaou-stained images.

In the cell clump classification stage, we design an ECRNet based on contrastive learning, which considers both instances and label facts. Specifically, different staining style images with the same classification label are considered similar. In addition, we learn meaningful and consistent anatomical features through the label contrastive loss, and introduce a label memory bank and a momentum update encoder to maintain encoded feature consistency. Experimental results show that our method achieves excellent performance on mixed staining style datasets, indirectly demonstrating its robustness. Compared to other methods, ECRNet achieves the best performance in both classification tasks with the two-stage strategy and the one-stage strategy.

Finally, there are two limitations of this work. On the one hand, because the data comes from a single institution, our approach is not externally validated on different institutional datasets. Although we have tried to ensure as much diversity as possible in the dataset during the data collection process, and used contrastive learning to enhance the generalization of the screening framework. However, we still lack external validation results from different medical centers. In future work, we will extend our method to other medical center datasets for external validation. In addition, annotation has been a challenge due to the scarce number of cytopathologists. We will focus on investigating self-supervised learning to reduce the annotation workload of cytopathologists.

Conclusions

In this paper, we present a clinically motivated deep learning framework for endometrial cancer cell clumps screening. in the first stage, we propose CM-UNet to obtain the ROI set, and the CA and MSS modules are able to fuse features from different scales to obtain more semantic information. In the second stage, we utilize ECRNet to classify ROIs. Contrastive learning is used to bring instances of the same class in the representation space closer together and push instances of different classes apart. Experiments show that our framework performs well on the XJTU-EC dataset. Our future work will focus on providing objective and complementary diagnostic input for clinical diagnosis, and supporting effective deployment by advanced algorithms. We believe that this can help reduce the burden on patients and physicians.

Data Availability

Data cannot be shared publicly because of privacy protection. The owner of the data is the Ethics Committee of the First Affiliated Hospital of Xi’an Jiaotong University, and therefore it is not freely available. The Ethics Committee of the First Affiliated Hospital of Xi’an Jiaotong University imposed these restrictions. Data are available from the Ethics Committee (xjyfyllh@163.com) for researchers who meet the criteria for access to confidential data. The externally validated data were obtained from the public data platform, AIstudio, accessible via the Internet (https://aistudio.baidu.com/datasetdetail/273988).Anyone can access this data after registering an account on this platform.

Funding Statement

This work is supported by National Natural Science Foundation of China (No. 62376211; 62206218), Natural Science Foundation of Zhejiang Province (No. LTGG23F030006), Special Project for Technological Innovation Guidance of Shaanxi Province (No. 2024ZC-YYDP-24).

References

  • 1.Urick ME, Bell DW. Clinical actionability of molecular targets in endometrial cancer. Nature Reviews Cancer. 2019;19(9):510–21. doi: 10.1038/s41568-019-0177-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Kong A, Johnson N, Kitchener HC, Lawrie TA. Adjuvant radiotherapy for stage I endometrial cancer. Cochrane Database Syst Rev. 2012;2012(4):Cd003916. Epub 2012/04/20. doi: 10.1002/14651858.CD003916.pub4 . [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Weiderpass E, Hashim D, Labrèche F. Malignant Tumors of the Female Reproductive System. In: Anttila S, Boffetta P, editors. Occupational Cancers. Cham: Springer International Publishing; 2020. p. 439–53. [Google Scholar]
  • 4.Lortet-Tieulent J, Ferlay J, Bray F, Jemal A. International patterns and trends in endometrial cancer incidence, 1978–2013. JNCI: Journal of the National Cancer Institute. 2018;110(4):354–61. doi: 10.1093/jnci/djx214 [DOI] [PubMed] [Google Scholar]
  • 5.Raglan O, Kalliala I, Markozannes G, Cividini S, Gunter MJ, Nautiyal J, et al. Risk factors for endometrial cancer: An umbrella review of the literature. International Journal of Cancer. 2019;145(7):1719–30. doi: 10.1002/ijc.31961 [DOI] [PubMed] [Google Scholar]
  • 6.Nishida N, Murakami F, Kuroda A, Sakamoto Y, Tasaki K, Tasaki S, et al. Clinical utility of endometrial cell block cytology in postmenopausal women. Acta Cytologica. 2017;61(6):441–6. doi: 10.1159/000479307 [DOI] [PubMed] [Google Scholar]
  • 7.Pei Z, Cao S, Lu L, Chen W. Direct Cellularity Estimation on Breast Cancer Histopathology Images Using Transfer Learning. Computational and Mathematical Methods in Medicine. 2019;2019:3041250. doi: 10.1155/2019/3041250 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Bao H-j, Chen X, Liu X, Wu W, Li Q-h, Xian J-y, et al. Box C/D snoRNA SNORD89 influences the occurrence and development of endometrial cancer through 2’-O-methylation modification of Bim. Cell Death Discovery. 2022;8(1):309. doi: 10.1038/s41420-022-01102-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Liu Y-S, Wang H-M, Gao Y. Controversy on Positive Peritoneal Cytology of Endometrial Carcinoma. Computational and Mathematical Methods in Medicine. 2022;2022:1906769. doi: 10.1155/2022/1906769 [DOI] [PMC free article] [PubMed] [Google Scholar] [Retracted]
  • 10.Kondo E, Tabata T, Koduka Y, Nishiura K, Tanida K, Okugawa T, et al. What is the best method of detecting endometrial cancer in outpatients?-endometrial sampling, suction curettage, endometrial cytology. Cytopathology. 2008;19(1):28–33. doi: 10.1111/j.1365-2303.2007.00509.x [DOI] [PubMed] [Google Scholar]
  • 11.Buccoliero AM, Gheri CF, Castiglione F, Garbini F, Barbetti A, Fambrini M, et al. Liquid-based endometrial cytology: cyto-histological correlation in a population of 917 women. Cytopathology. 2007;18(4):241–9. doi: 10.1111/j.1365-2303.2007.00463.x [DOI] [PubMed] [Google Scholar]
  • 12.Wadhwa N, Jatawa SK, Tiwari A. Non-invasive urine based tests for the detection of bladder cancer. Journal of clinical pathology. 2012;65(11):970–5. doi: 10.1136/jclinpath-2012-200812 [DOI] [PubMed] [Google Scholar]
  • 13.Sawaya GF, Smith-McCune K, Kuppermann M. Cervical cancer screening: more choices in 2019. Jama. 2019;321(20):2018–9. doi: 10.1001/jama.2019.4595 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.O’Flynn H, Ryan NAJ, Narine N, Shelton D, Rana D, Crosbie EJ. Diagnostic accuracy of cytology for the detection of endometrial cancer in urine and vaginal samples. Nature Communications. 2021;12(1):952. doi: 10.1038/s41467-021-21257-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Wang Q, Wang Q, Zhao L, Han L, Sun C, Ma S, et al. Endometrial Cytology as a Method to Improve the Accuracy of Diagnosis of Endometrial Cancer: Case Report and Meta-Analysis. Frontiers in Oncology. 2019;9. doi: 10.3389/fonc.2019.00256 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Ke J, Shen Y, Lu Y, Deng J, Wright JD, Zhang Y, et al. Quantitative analysis of abnormalities in gynecologic cytopathology with deep learning. Laboratory Investigation. 2021;101(4):513–24. doi: 10.1038/s41374-021-00537-1 [DOI] [PubMed] [Google Scholar]
  • 17.Landau MS, Pantanowitz L. Artificial intelligence in cytopathology: a review of the literature and overview of commercial landscape. Journal of the American Society of Cytopathology. 2019;8(4):230–41. doi: 10.1016/j.jasc.2019.03.003 [DOI] [PubMed] [Google Scholar]
  • 18.Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542(7639):115–8. doi: 10.1038/nature21056 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Lilli L, Giarnieri E, Scardapane S. A Calibrated Multiexit Neural Network for Detecting Urothelial Cancer Cells. Computational and Mathematical Methods in Medicine. 2021;2021:5569458. doi: 10.1155/2021/5569458 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Dov D, Kovalsky SZ, Assaad S, Cohen J, Range DE, Pendse AA, et al. Weakly supervised instance learning for thyroid malignancy prediction from whole slide cytopathology images. Medical Image Analysis. 2021;67:101814. doi: 10.1016/j.media.2020.101814 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Ma J, Yu J, Liu S, Chen L, Li X, Feng J, et al. PathSRGAN: Multi-Supervised Super-Resolution for Cytopathological Images Using Generative Adversarial Network. IEEE Transactions on Medical Imaging. 2020;39(9):2920–30. doi: 10.1109/TMI.2020.2980839 [DOI] [PubMed] [Google Scholar]
  • 22.Wei J, Ji Q, Gao Y, Yang X, Guo D, Gu D, et al. A multi-scale, multi-region and attention mechanism-based deep learning framework for prediction of grading in hepatocellular carcinoma. Medical Physics. n/a(n/a). doi: 10.1002/mp.16127 [DOI] [PubMed] [Google Scholar]
  • 23.Sun SW, Xu X, Liu QP, Chen JN, Zhu FP, Liu XS, et al. LiSNet: An artificial intelligence -based tool for liver imaging staging of hepatocellular carcinoma aggressiveness. Medical Physics. 2022;49(11):6903–13. doi: 10.1002/mp.15972 [DOI] [PubMed] [Google Scholar]
  • 24.Zhao Q, He Y, Wu Y, Huang D, Wang Y, Sun C, et al. Vocal cord lesions classification based on deep convolutional neural network and transfer learning. Medical Physics. 2022;49(1):432–42. doi: 10.1002/mp.15371 [DOI] [PubMed] [Google Scholar]
  • 25.Hodneland E, Dybvik JA, Wagner-Larsen KS, Šoltészová V, Munthe-Kaas AZ, Fasmer KE, et al. Automated segmentation of endometrial cancer on MR images using deep learning. Scientific Reports. 2021;11(1):179. doi: 10.1038/s41598-020-80068-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Xia Z, Zhang L, Liu S, Ran W, Liu Y, Tu J. Deep learning-based hysteroscopic intelligent examination and ultrasound examination for diagnosis of endometrial carcinoma. The Journal of Supercomputing. 2022;78(9):11229–44. doi: 10.1007/s11227-021-04046-2 [DOI] [Google Scholar]
  • 27.Xiong L, Chen C, Lin Y, Mao W, Song Z. A computer-aided determining method for the myometrial infiltration depth of early endometrial cancer on MRI images. BioMedical Engineering OnLine. 2023;22(1):103. doi: 10.1186/s12938-023-01169-w [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Otani S, Himoto Y, Nishio M, Fujimoto K, Moribata Y, Yakami M, et al. Radiomic machine learning for pretreatment assessment of prognostic risk factors for endometrial cancer and its effects on radiologists’ decisions of deep myometrial invasion. Magnetic Resonance Imaging. 2022;85:161–7. doi: 10.1016/j.mri.2021.10.024 [DOI] [PubMed] [Google Scholar]
  • 29.Mao W, Chen C, Gao H, Xiong L, Lin Y. Quantitative evaluation of myometrial infiltration depth ratio for early endometrial cancer based on deep learning. Biomedical Signal Processing and Control. 2023;84:104685. doi: 10.1016/j.bspc.2023.104685 [DOI] [Google Scholar]
  • 30.Dong H-C, Dong H-K, Yu M-H, Lin Y-H, Chang C-C. Using Deep Learning with Convolutional Neural Network Approach to Identify the Invasion Depth of Endometrial Cancer in Myometrium Using MR Images: A Pilot Study. International Journal of Environmental Research and Public Health. 2020;17(16):5993. doi: 10.3390/ijerph17165993 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Fell C, Mohammadi M, Morrison D, Arandjelović O, Syed S, Konanahalli P, et al. Detection of malignancy in whole slide images of endometrial cancer biopsies using artificial intelligence. Plos one. 2023;18(3):e0282577. doi: 10.1371/journal.pone.0282577 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Fremond S, Andani S, Wolf JB, Dijkstra J, Melsbach S, Jobsen JJ, et al. Interpretable deep learning model to predict the molecular classification of endometrial cancer from haematoxylin and eosin-stained whole-slide images: a combined analysis of the PORTEC randomised trials and clinical cohorts. The Lancet Digital Health. 2023;5(2):e71–e82. doi: 10.1016/S2589-7500(22)00210-2 [DOI] [PubMed] [Google Scholar]
  • 33.Feng M, Zhao Y, Chen J, Zhao T, Mei J, Fan Y, et al. A deep learning model for lymph node metastasis prediction based on digital histopathological images of primary endometrial cancer. Quant Imaging Med Surg. 2023;13(3):1899–913. Epub 2023/03/15. doi: 10.21037/qims-22-220 . [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Kurata Y, Nishio M, Moribata Y, Kido A, Himoto Y, Otani S, et al. Automatic segmentation of uterine endometrial cancer on multi-sequence MRI using a convolutional neural network. Scientific Reports. 2021;11(1):14440. doi: 10.1038/s41598-021-93792-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Urushibara A, Saida T, Mori K, Ishiguro T, Inoue K, Masumoto T, et al. The efficacy of deep learning models in the diagnosis of endometrial cancer using MRI: a comparison with radiologists. BMC Medical Imaging. 2022;22(1):80. doi: 10.1186/s12880-022-00808-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Zhang Y, Chen S, Wang Y, Li J, Xu K, Chen J, et al. Deep learning-based methods for classification of microsatellite instability in endometrial cancer from HE-stained pathological images. Journal of Cancer Research and Clinical Oncology. 2023;149(11):8877–88. doi: 10.1007/s00432-023-04838-4 [DOI] [PubMed] [Google Scholar]
  • 37.Zhang X, Ba W, Zhao X, Wang C, Li Q, Zhang Y, et al. Clinical-grade endometrial cancer detection system via whole-slide images using deep learning. Frontiers in Oncology. 2022;12. doi: 10.3389/fonc.2022.1040238 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38.Hong R, Liu W, DeLair D, Razavian N, Fenyö D. Predicting endometrial cancer subtypes and molecular features from histopathology images using multi-resolution deep learning models. Cell Reports Medicine. 2021;2(9):100400. doi: 10.1016/j.xcrm.2021.100400 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Lin H, Chen H, Wang X, Wang Q, Wang L, Heng P-A. Dual-path network with synergistic grouping loss and evidence driven risk stratification for whole slide cervical image analysis. Medical Image Analysis. 2021;69:101955. doi: 10.1016/j.media.2021.101955 [DOI] [PubMed] [Google Scholar]
  • 40.Shaban MT, Baur C, Navab N, Albarqouni S, editors. Staingan: Stain Style Transfer for Digital Histological Images. 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019); 2019 8–11 April 2019.
  • 41.Geng X, Liu X, Cheng S, Zeng S. Cervical cytopathology image refocusing via multi-scale attention features and domain normalization. Medical Image Analysis. 2022;81:102566. doi: 10.1016/j.media.2022.102566 [DOI] [PubMed] [Google Scholar]
  • 42.Maksem JA, Meiers I, Robboy SJ. A primer of endometrial cytology with histological correlation. Diagnostic cytopathology. 2007;35(12):817–44. doi: 10.1002/dc.20745 [DOI] [PubMed] [Google Scholar]
  • 43.Association WM. World Medical Association Declaration of Helsinki: ethical principles for medical research involving human subjects. Jama. 2013;310(20):2191–4. doi: 10.1001/jama.2013.281053 [DOI] [PubMed] [Google Scholar]
  • 44.Ronneberger O, Fischer P, Brox T, editors. U-net: Convolutional networks for biomedical image segmentation. International Conference on Medical image computing and computer-assisted intervention; 2015: Springer.
  • 45.Zhou Z, Siddiquee MMR, Tajbakhsh N, Liang J. UNet++: Redesigning Skip Connections to Exploit Multiscale Features in Image Segmentation. IEEE Transactions on Medical Imaging. 2020;39(6):1856–67. doi: 10.1109/TMI.2019.2959609 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46.Khan SD, Alarabi L, Basalamah S. An Encoder–Decoder Deep Learning Framework for Building Footprints Extraction from Aerial Imagery. Arabian Journal for Science and Engineering. 2023;48(2):1273–84. doi: 10.1007/s13369-022-06768-8 [DOI] [Google Scholar]
  • 47.Liu J, Guo X, Li B, Yuan Y, editors. COINet: Adaptive Segmentation with Co-Interactive Network for Autonomous Driving. 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS); 2021 27 Sept.-1 Oct. 2021.
  • 48.Fu J, Liu J, Tian H, Li Y, Bao Y, Fang Z, et al., editors. Dual Attention Network for Scene Segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 2019 15–20 June 2019.
  • 49.Wu H, Wang W, Zhong J, Lei B, Wen Z, Qin J. SCS-Net: A Scale and Context Sensitive Network for Retinal Vessel Segmentation. Medical Image Analysis. 2021;70:102025. doi: 10.1016/j.media.2021.102025 [DOI] [PubMed] [Google Scholar]
  • 50.Ren Y, Bai J, Zhang J, editors. Label Contrastive Coding Based Graph Neural Network for Graph Classification. Database Systems for Advanced Applications; 2021. 2021//; Cham: Springer International Publishing. [Google Scholar]
  • 51.He K, Fan H, Wu Y, Xie S, Girshick R, editors. Momentum Contrast for Unsupervised Visual Representation Learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 2020 13–19 June 2020.
  • 52.Woo S, Park J, Lee J-Y, Kweon IS, editors. CBAM: Convolutional Block Attention Module. Computer Vision–ECCV 2018; 2018 2018//; Cham: Springer International Publishing. [Google Scholar]
  • 53.Liu S, Deng W, editors. Very deep convolutional neural network based image classification using small training sample size. 2015 3rd IAPR Asian Conference on Pattern Recognition (ACPR); 2015 3–6 Nov. 2015.
  • 54.Dice LR. Measures of the Amount of Ecologic Association Between Species. Ecology. 1945;26(3):297–302. doi: 10.2307/1932409 [DOI] [Google Scholar]
  • 55.Kingma DP, Ba J. Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980. 2014.
  • 56.Shelhamer E, Long J, Darrell T. Fully Convolutional Networks for Semantic Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2017;39(4):640–51. doi: 10.1109/TPAMI.2016.2572683 [DOI] [PubMed] [Google Scholar]
  • 57.Chaurasia A, Culurciello E, editors. LinkNet: Exploiting encoder representations for efficient semantic segmentation. 2017 IEEE Visual Communications and Image Processing (VCIP); 2017 10–13 Dec. 2017.
  • 58.Chen L-C, Papandreou G, Schroff F, Adam H. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:170605587. 2017.
  • 59.Chen L-C, Zhu Y, Papandreou G, Schroff F, Adam H, editors. Encoder-decoder with atrous separable convolution for semantic image segmentation. Proceedings of the European conference on computer vision (ECCV); 2018.
  • 60.Howard AG, Zhu M, Chen B, Kalenichenko D, Wang W, Weyand T, et al. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:170404861. 2017.
  • 61.He K, Zhang X, Ren S, Sun J, editors. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition; 2016.
  • 62.Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z, editors. Rethinking the inception architecture for computer vision. Proceedings of the IEEE conference on computer vision and pattern recognition; 2016.
  • 63.Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, Unterthiner T, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:201011929. 2020.
  • 64.Xie S, Girshick R, Dollár P, Tu Z, He K, editors. Aggregated Residual Transformations for Deep Neural Networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2017 21–26 July 2017.
  • 65.Tan M, Le Q, editors. Efficientnet: Rethinking model scaling for convolutional neural networks. International conference on machine learning; 2019: PMLR.
  • 66.Huang G, Liu Z, Van Der Maaten L, Weinberger KQ, editors. Densely connected convolutional networks. Proceedings of the IEEE conference on computer vision and pattern recognition; 2017.
  • 67.Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:14091556. 2014.
  • 68.Xu Y, Jia Z, Wang LB, Ai Y, Zhang F, Lai M, et al. Large scale tissue histopathology image classification, segmentation, and visualization via deep convolutional activation features. BMC Bioinformatics. 2017;18(1):281. Epub 2017/05/28. doi: 10.1186/s12859-017-1685-x . [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 69.Endometrial Cancer; 2024 [cited 2024 Jun 4]. Database: AIstudio [Internet]. https://aistudio.baidu.com/datasetdetail/273988.

Decision Letter 0

Steve Zimmerman

7 Dec 2023

PONE-D-23-24291Deep Learning-based Endometrial Cancer Screening System for Multimodal Cytopathology ImagePLOS ONE

Dear Dr. Wang,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

The manuscript has been evaluated by three reviewers, and their comments are available below.

The reviewers have raised a number of concerns that need attention. They request additional information and clarification in several areas of the manuscript, including methodological aspects of the study

Could you please revise the manuscript to carefully address the concerns raised?

Please submit your revised manuscript by Jan 21 2024 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

  • A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

  • A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

  • An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Steve Zimmerman, PhD

Senior Editor, PLOS ONE

Journal requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. During our evaluation of the documents provided, we noted that your ethics approval letter did not cover the entire date range for participant recruitment. Before we can proceed further with the submission, please provide the ethics approval extension documents, along with English translations.

3. Thank you for stating in your Funding Statement:

“This work was supported in part by the National Natural Science Foundation of China

under Grant 62206218, in part by Natural Science Foundation of Zhejiang Province

under Grant LTGG23F030006. Grantees have helped in study design, data collection

and analysis, publication decisions or manuscript preparation.”

Please provide an amended statement that declares *all* the funding or sources of support (whether external or internal to your organization) received during this study, as detailed online in our guide for authors at http://journals.plos.org/plosone/s/submit-now.  Please also include the statement “There was no additional external funding received for this study.” in your updated Funding Statement.

Please include your amended Funding Statement within your cover letter. We will change the online submission form on your behalf.

4. We note that you have indicated that data from this study are available upon request. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For more information on unacceptable data access restrictions, please see http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data-access-restrictions.

In your revised cover letter, please address the following prompts:

a) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially sensitive information, data are owned by a third-party organization, etc.) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent.

b) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings as either Supporting Information files or to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. For a list of acceptable repositories, please see http://journals.plos.org/plosone/s/data-availability#loc-recommended-repositories.

We will update your Data Availability statement on your behalf to reflect the information you provide.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Title : Deep Learning-based Endometrial Cancer Screening System for Multimodal Cytopathology Image

Summary/Contribution: The paper addresses a critical and clinically relevant issue in the field of cancer screening. The development of an automated framework for endometrial cancer detection is commendable, considering the significant impact it can have on early diagnosis and patient outcomes. It is noteworthy that the authors have invested the time and resources to develop the XJTU-EC dataset, which provides a valuable resource for this field.

Strengths:

• The authors' dedication to collecting a comprehensive dataset over several years is a significant contribution to the field. The XJTU-EC dataset is expected to facilitate further research in endometrial cancer screening.

• The paper explains the proposed two-stage framework clearly, providing insights into CM-Unet and ECRNet, which are crucial for understanding the methodology. ECRNet uses attention modules and label memory banks to add depth to the methodology. The paper reports that the proposed framework outperforms other classical models.

• The framework has the potential to greatly benefit cytopathologists by enhancing their efficiency. The paper also emphasizes the importance of detection, in cancer and addresses a practical requirement.

• The authors are transparent about the ethical approval process, which adds credibility to the research.

Major Comments:

• How representative is the XJTU-EC dataset of endometrial cytopathology images, considering it was collected over several years from a single institution? How does it account for potential bias in the dataset? The paper should discuss the generalization potential of the proposed framework to datasets from other medical centers and possibly provide results or insights from external validation on diverse datasets.

• The paper mentions that different medical centers use different staining styles. How robust is the proposed framework to handle extreme variations in staining styles, and can it generalize to datasets from other institutions?

• Could you elaborate on the inter-pathologist agreement in annotating cell clumps? How did you address any disagreements in labeling during the annotation process? What measures were taken to ensure the accuracy and consistency of the segmentation and classification labels given the complexity of cytological images?

• Has there been any validation or real world implementation to evaluate how the proposed framework has affected patient outcomes? Have cytopathologists been engaged in assessing its effectiveness in a clinical setting.

Further Comments: In addition to the areas for improvements mentioned above, the authors should also consider the following matters:

• There are some type in the write-up such as in Introduction part “ negative exsample pairs”- which should be revised.

• The paper lacks figure numbers, making it challenging to correlate specific references in the text to corresponding figures.

Reviewer #2: The authors do deep learning on a set of endometrial lesions based on a variety of available CNN modalities along with their own modality. Their cases were annotated by pathologists and many The data highlights the strengths and weaknesses of currently available algorithms. Their detailed methods could serve as a template to validate other cytology specimens. The code could be validated on larger, institutional and multiinstitutional datasets in the future as it becomes publicly available.

Several questions

You annotate many different endometrial cancers. Could you provide more information about the tumor types on follow up such as endometrioid, endocervical, mucinous, high grade serous etc? as a pathologist I could have a better understanding of how the cases were annotated and what the algorithm would be evaluating if I knew the different tumor types included

similarly for benign is there a histologic follow up?

Reviewer #3: 1) Title:

The title is short and not clear.

2) Abstract:

a) The abstract is too short and it seems as a preface to a report.

b) The objectives of this research are not clear. The authors need to explain exactly what the aim of this research.

3) Introduction:

a) In the introduction section, Authors should produce distinctive features mentioning their unique interpretation of the research along with comparative analysis of existing techniques to strengthen their work. Authors can focus on referring more recent journal papers to include further study.

b) There were some linguistic errors and typos.

4) References:

a) The number of references seems sufficient, but most of the references are not recent.

b) References are not well distributed.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: Yes: Jordan P. Reynolds MD

Reviewer #3: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

Attachment

Submitted filename: review paper.docx

pone.0306549.s001.docx (14.9KB, docx)
PLoS One. 2024 Jul 31;19(7):e0306549. doi: 10.1371/journal.pone.0306549.r002

Author response to Decision Letter 0


18 Jan 2024

Dear Editor and Reviewers,

Thank you very much for your careful review of our manuscript " A Deep Learning Framework for Predicting Endometrial Cancer from Cytopathologic Images with Different Staining Styles" (ID: PONE-D-23-24291). We have revised the manuscript thoroughly according to your comments and suggestions. These main changes in the revised manuscript are highlighted in red. The revisions for grammatical errors and improved sentences are marked in blue. Detailed responses to your comments are provided below.

P.S.: Citations are numbered sequentially in the order in which they appear in the text.

Sincerely,

Dexing Zhong

Academic editor:

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf.

Responses:

Thank you for your reminder. We have revised the manuscript with reference to the PLOS ONE style template. These main changes in the revised manuscript are highlighted in red. The revisions for grammatical errors and improved sentences are marked in blue. We have determined that the manuscript meets PLOS ONE style requirements, including file naming requirements.

2. During our evaluation of the documents provided, we noted that your ethics approval letter did not cover the entire date range for participant recruitment. Before we can proceed further with the submission, please provide the ethics approval extension documents, along with English translations.

Responses:

Thank you for your suggestion. Our data were provided by the First Affiliated Hospital of Xi'an Jiaotong University. We apologize for the miscommunication with the data provider that led us to mistake the exact time of data collection. We reconfirmed with the data provider and clarified that the data collection lasted for one year, from November 2019 to 2020, which is covered by the ethical approval. After 2020 and up to 2022, two experienced cytopathologists annotated the data. We have revised the relevant parts of the manuscript. Thank you very much for your understanding.

3. Thank you for stating in your Funding Statement:

“This work was supported in part by the National Natural Science Foundation of China

under Grant 62206218, in part by Natural Science Foundation of Zhejiang Province

under Grant LTGG23F030006. Grantees have helped in study design, data collection

and analysis, publication decisions or manuscript preparation.”

Please provide an amended statement that declares *all* the funding or sources of support (whether external or internal to your organization) received during this study, as detailed online in our guide for authors at http://journals.plos.org/plosone/s/submit-now. Please also include the statement “There was no additional external funding received for this study.” in your updated Funding Statement.

Please include your amended Funding Statement within your cover letter. We will change the online submission form on your behalf.

Responses:

Thank you for your suggestion, we have included the Funding Statement at the end of the revised cover letter, which is copied in the attachment below.

Funding Statement:

This work was supported in part by the Clinical Research Award of the First Affiliated Hospital of Xi’an Jiaotong University (No. XJTU1AF-CRF-2019-002), in part by the National Natural Science Foundation of China (No. 62206218), in part by the Natural Science Foundation of Zhejiang Province (No. LTGG23F030006). Grantees have helped in study design, data collection and analysis, publication decisions or manuscript preparation.

There was no additional external funding received for this study.

4. We note that you have indicated that data from this study are available upon request. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For more information on unacceptable data access restrictions, please see http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data-access-restrictions.

In your revised cover letter, please address the following prompts:

a) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially sensitive information, data are owned by a third-party organization, etc.) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent.

b) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings as either Supporting Information files or to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. For a list of acceptable repositories, please see http://journals.plos.org/plosone/s/data-availability#loc-recommended-repositories.

We will update your Data Availability statement on your behalf to reflect the information you provide.

Responses:

Data cannot be shared publicly because it is patient data and the approvals do not permit us to release the data. Our data are owned by the First Affiliated Hospital of Xi'an Jiaotong University due to ethical committee requirements. If researchers need data, please send a data request email to xjyfyllh@163.com, which is the contact information for the Ethics Committee of the First Affiliated Hospital of Xi'an Jiaotong University.

Response to Reviewer #1:

The paper addresses a critical and clinically relevant issue in the field of cancer screening. The development of an automated framework for endometrial cancer detection is commendable, considering the significant impact it can have on early diagnosis and patient outcomes. It is noteworthy that the authors have invested the time and resources to develop the XJTU-EC dataset, which provides a valuable resource for this field.

Responses:

At first, we would like to sincerely thank the anonymous reviewers for your time and concerns to our manuscript, especially your helpful comments and suggestions to improve the quality of our paper. With your kindly help, we have carefully revised the entire manuscript. All the revisions have been clearly highlighted which are visible to the editors and reviewers.

Major comments:

1. How representative is the XJTU-EC dataset of endometrial cytopathology images, considering it was collected over several years from a single institution? How does it account for potential bias in the dataset? The paper should discuss the generalization potential of the proposed framework to datasets from other medical centers and possibly provide results or insights from external validation on diverse datasets.

Response:

We thank the reviewer for the suggestion. In response, we have added a description of the representativeness of our dataset in the Discussion section. Since our dataset contains both papanicolaou and H&E stained images, which are the most common staining modalities for cytology images. Therefore, it can be somewhat considered a representative dataset. In addition, the data were collected across patients of different ages, so the XJTU-EC dataset has more diversity. And the diversity of the training data can help the neural network overcome the effects of potential bias. Our method achieves excellent performance on this dataset, which can demonstrate its generalization ability. We will be collecting more data in the future to further minimize potential bias in the dataset. And we are committed to generalizing the framework presented in this paper to other medical center datasets, and related external validation experiments are in progress. This is part of our future work. The detailed changes (highlighted in red color) can be seen in Discussion in the revised manuscript and the following attachment.

Discussion (page 19-20, line 404-413):

Due to the lack of relevant data and the complexity of cell morphology, endometrial cancer screening based on cytopathology is difficult to promote. Therefore, our team collected and annotated WSIs from 139 patients to create the first endometrial cytopathology image dataset, named the XJTU-EC dataset. Since our dataset contains both papanicolaou and H&E stained images, which are the most common staining modalities for cytopathology images. Therefore, it can be somewhat considered a representative dataset. In addition, the data includes patients of different age, so the XJTU-EC dataset has more diversity. Based on this dataset, we investigated the first clinically automated deep learning framework for extracting and identifying normal or cancerous endometrial cell clumps. The results will be presented to cytopathologists as a reference.

Discussion (page 20-21, line 429-437):

Finally, there are two limitations of this work. On the one hand, because the data comes from a single institution, our approach is not externally validated on different institutional datasets. Although we have tried to ensure as much diversity as possible in the dataset during the data collection process, and used contrastive learning to enhance the generalization of the screening framework. However, we still lack external validation results from different medical centers. In future work, we will extend our method to other medical center datasets. In addition, annotation has been a challenge due to the scarce number of cytopathologists. We will focus on investigating self-supervised learning to reduce the annotation workload of cytopathologists.

2. The paper mentions that different medical centers use different staining styles. How robust is the proposed framework to handle extreme variations in staining styles, and can it generalize to datasets from other institutions?

Response:

Special thanks to you for your good comments. To the best of our knowledge, papanicolaou staining and H&E staining are the most commonly used staining styles for cytology images. In order to adapt to these two different staining methods, in the cell clump extraction stage, we use the robust Unet as the backbone, which has been previously generalized to many datasets. Based on this, we introduced the CA module and MSS module to further improve the target segmentation accuracy. The segmentation results show that CM-Unet can well ignore the effect of staining style and accurately extract ROIs.

In the cell clump classification stage, we design an ECRNet based on contrastive learning, which considers both instances and label facts. Specifically, different staining style images with the same classification label are considered similar. Thus, it is able to adapt to datasets with changing staining styles. In addition, we learn meaningful and consistent anatomical features through the label contrastive loss Lcon, and introduce a label memory bank and a momentum update encoder to maintain encoded feature consistency. Experimental results show that our method achieves excellent performance on mixed staining style datasets, indirectly demonstrating its robustness. We are actively promoting our method to datasets from other institutions, which has been written into our future work. The detailed changes (highlighted in red color) can be seen in Discussion in the revised manuscript and the following attachment.

Discussion (page 20, line 414-428):

In order to adapt to different staining styles, in the cell clump extraction stage, we use the robust Unet as the backbone, which has been previously generalized to many datasets. Based on this, we introduced the CA module pay attention to global contextual information, and MSS module to aggregate semantic features at multiple scales. So, our method achieves better segmentation results. Experiments demonstrate that CM-Unet is able to perform well on both H&E-stained images and papanicolaou-stained images.

In the cell clump classification stage, we design an ECRNet based on contrastive learning, which considers both instances and label facts. Specifically, different staining style images with the same classification label are considered similar. Thus, it is able to adapt to datasets with changing staining styles. In addition, we learn meaningful and consistent anatomical features through the label contrastive loss, and introduce a label memory bank and a momentum update encoder to maintain encoded feature consistency. Experimental results show that our method achieves excellent performance on mixed staining style datasets, indirectly demonstrating its robustness. Compared to other methods, ECRNet achieves the best performance in both classification tasks with the two-stage strategy and the one-stage strategy.

3. Could you elaborate on the inter-pathologist agreement in annotating cell clumps? How did you address any disagreements in labeling during the annotation process? What measures were taken to ensure the accuracy and consistency of the segmentation and classification labels given the complexity of cytological images?

Response:

Thank you for your time and patience. We have added a description of the data annotation process in the Methods section. First, a senior cytopathologist annotates the images for segmentation labels, which are reviewed by another cytopathologist. After the review is accurate, we move on to the next process. Undoubtedly, the cell clumps in the negative slides are all negative, but the ones in the positive slides have both negative and positive cell clumps. Therefore, the two cytopathologists again vote on the labeling of each cell clump, when the votes do not agree, they will discuss it. If the discussion fails to result in an accurate diagnosis, the cell clump is discarded. These measures ensure the accuracy and consistency of segmentation and classification labels. The detailed changes (highlighted in red color) can be seen in Methods in the revised manuscript and the following attachment.

Methods (page 8-9, line 173-192):

And the image annotation process includes segmentation label annotation and classification label annotation. Segmentation labels were obtained by two experienced pathologists using Adobe Photoshop CC (2019 v20.0.2.30). First, one senior cytopathologist segmented the cell clumps and the results were reviewed by the other cytopathologist. After the review is accurate, the pathologists annotated the cell clumps according to the International Society of Gynecologic Pathologists and the 2014 World Health Organization classification of uterine tumors. All cell clumps were classified into two categories: malignant (atypical cells of undetermined significance, suspected malignant tumor cells, and malignant tumor cells), and benign (non-malignant tumor cells). Benign diagnosis is defined as cell clumps with neat edges, nuclei with oval or spindle shape, and evenly distributed, finely granular chromatin. Malignant diagnosis referred to a three-dimensional appearance, irregular (including dilated, branched, protruding, and papillotubular) edge, with the nucleus poloidal disordering or disappearing (including megakaryocyte appearance, nuclear membrane thickness, and coarse granular or coarse block chromatin). Both benign and malignant tumors were followed up histologically. Undoubtedly, the cell clumps in the negative slides are all negative, but the ones in the positive slides have both negative and positive cell clumps. Therefore, the two cytopathologists again vote on the labeling of each cell clump, when the votes do not agree, they will discuss it. If the discussion fails to result in an accurate diagnosis, the cell clump is discarded. These measures ensure the accuracy and consistency of segmentation and classification labels.

4. Has there been any validation or real world implementation to evaluate how the proposed framework has affected patient outcomes? Have cytopathologists been engaged in assessing its effectiveness in a clinical setting.

Response:

We thank the reviewer for the suggestion. We have two cytopathologists who have been evaluating the effectiveness of the framework in actual clinical applications. After evaluating

Attachment

Submitted filename: Response to Reviewers.docx

pone.0306549.s002.docx (51.6KB, docx)

Decision Letter 1

Kazunori Nagasaka

28 Feb 2024

PONE-D-23-24291R1A Deep Learning Framework for Predicting Endometrial Cancer from Cytopathologic Images with Different Staining StylesPLOS ONE

Dear Dr. Wang,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Please submit your revised manuscript by Apr 13 2024 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

  • A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

  • A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

  • An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Kazunori Nagasaka

Academic Editor

PLOS ONE

Additional Editor Comments:

Dear Authors,

Still the explanation for the database is lacking and more detailed information for expertimental procedure is needed.

Please provide these information further in the text.

The manuscript will not be accepted as current stand and the authors should answer the reviewers inquries point by point.

Sincerely,

Kazunori Nagasaka

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #4: (No Response)

Reviewer #5: All comments have been addressed

Reviewer #6: (No Response)

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #4: No

Reviewer #5: Yes

Reviewer #6: Partly

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #4: N/A

Reviewer #5: Yes

Reviewer #6: N/A

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #4: Yes

Reviewer #5: Yes

Reviewer #6: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #4: Yes

Reviewer #5: Yes

Reviewer #6: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Thank you for providing your responses. I have reviewed the answers, and they satisfy my inquiries. I don't have any further comments at this time.

Reviewer #4: The paper lacks sufficient straightforward considerations to be accepted, as highlighted by Major comment 1 of Reviewer 1 and the corresponding response.

Major issue 1:

Reviewer 1 and I share concerns about the potential bias in the XJTU-EC dataset. It is challenging to accept AI research that lacks transparent and adequate teacher data. The authors have stated that additional external validation experiments are being conducted, and until then, this paper will not reach the level of acceptance. The authors consider this as future work, but it is the most crucial task that needs to be done now. Without validation in an external cohort, only a limited number of articles on AI-assisted cytology can be accepted. In 2024, it is not feasible to claim that an AI has been developed solely based on data from a single institution.

Major issue 2:

It appears that even after the revision, the description of the XJTU dataset is still inadequate. I understand that there are too few cases to start with. I am unsure about what was defined as "the first large cytopathology image dataset". For reference, the dataset used by Kanavati et al. to develop AI for cervical cytology only had 3121 cases with Papanicolaou staining [PMID: 35267466]. It is not surprising to see different ages included in the study as collecting 139 cases of the same age would be difficult. It is important to provide specific details about the age dispersion to maintain transparency. It's concerning that 139 cases lack mention of lesions or histology. Could you please provide information about their pregnancy and childbirth history?  Whether or not age alone creates bias in data from a single center is not clear; exclusion criteria inconsistencies exist, such as allowing ovarian cancer but excluding cervical cancer.

Reviewer #5: The channel attention module is similar with the scale-aware distilled decoder proposed in ''COINet: Adaptive Segmentation with Co-Interactive Network for Autonomous Driving''. This paper should be cited.

Liu, Jie, et al. "COINet: Adaptive Segmentation with Co-Interactive Network for Autonomous Driving." 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021.

Reviewer #6: The paper a framework for endometrial cancer using deep learning, addressing the challenge of analyzing cytopathology images with different staining styles. The framework includes a novel CM-UNet for cell clump segmentation and an ECRNet classification algorithm based on contrastive learning.

1. The contribution of proposed work is missing. The authors need to highlight the research gaps and should explicitly mention how proposed framework filled those gaps.

2. In cytopathological images, cells can exhibit diverse shapes and sizes, making it difficult for an algorithm to generalize well across different cell types and imaging conditions. How the proposed method addresses this problem?

3. Please include one figure that illustrates the overall flow of the framework. From input image to segmentation output.

4. Please include high resolution images

5. Since the framework is solving the segmentation problem, therefore, following references need to discussed in the revised manuscript.

a. An Encoder–Decoder Deep Learning Framework for Building Footprints Extraction from Aerial Imagery. Arabian Journal for Science and Engineering, pp.1-12.

6. How many parameters are used to train the network?

7. The details of loss function are missing.

8. Experiment section is very weak. The authors need to provide more quantitative and qualitative evaluation of the framework as well as comparisons with other reference methods.

9. Please discuss failure cases and also provide their justifications

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #4: No

Reviewer #5: No

Reviewer #6: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

PLoS One. 2024 Jul 31;19(7):e0306549. doi: 10.1371/journal.pone.0306549.r004

Author response to Decision Letter 1


19 Apr 2024

Dear Editor and Reviewers,

Thank you very much for your careful review of our manuscript " A Deep Learning Framework for Predicting Endometrial Cancer from Cytopathologic Images with Different Staining Styles" (ID: PONE-D-23-24291). We have revised the manuscript thoroughly according to your comments and suggestions. These main changes in the revised manuscript are highlighted in red. The revisions for grammatical errors and improved sentences are marked in blue. Detailed responses to your comments are provided below.

P.S.: Citations are numbered sequentially in the order in which they appear in the text. Subject to formatting and typographical constraints, etc., a more detailed response can be found in the attached attachment.

Sincerely,

Dexing Zhong

Attachment

Submitted filename: Response to Reviewers.docx

pone.0306549.s003.docx (735.8KB, docx)

Decision Letter 2

Kazunori Nagasaka

21 May 2024

PONE-D-23-24291R2A Deep Learning Framework for Predicting Endometrial Cancer from Cytopathologic Images with Different Staining StylesPLOS ONE

Dear Dr. Wang,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Please submit your revised manuscript by Jul 05 2024 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

  • A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

  • A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

  • An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Kazunori Nagasaka

Academic Editor

PLOS ONE

Additional Editor Comments:

Dear Authors,

Please reply to Reviwers 4 comment.

Sincerely,

Kazunori Nagasaka

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #4: (No Response)

Reviewer #5: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #4: No

Reviewer #5: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #4: N/A

Reviewer #5: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #4: No

Reviewer #5: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #4: (No Response)

Reviewer #5: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #4: The author was unable to respond to the comments made by reviewer #4. In 2024, medical machine learning research with a small number of data as teachers and without external validation should not be considered as a research paper, or at the very least, should not be published in PLOS One.

Reviewer #5: (No Response)

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #4: No

Reviewer #5: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

PLoS One. 2024 Jul 31;19(7):e0306549. doi: 10.1371/journal.pone.0306549.r006

Author response to Decision Letter 2


4 Jun 2024

Dear Editor and Reviewers,

Thank you very much for your careful review of our manuscript " A Deep Learning Framework for Predicting Endometrial Cancer from Cytopathologic Images with Different Staining Styles" (ID: PONE-D-23-24291). We have revised the manuscript thoroughly according to your comments and suggestions. These main changes in the revised manuscript are highlighted in red. The revisions for grammatical errors and improved sentences are marked in blue. Detailed responses to your comments are provided below.

P.S.: Citations are numbered sequentially in the order in which they appear in the text.

Sincerely,

Dexing Zhong

Response to Reviewer #4:

The author was unable to respond to the comments made by reviewer #4. In 2024, medical machine learning research with a small number of data as teachers and without external validation should not be considered as a research paper, or at the very least, should not be published in PLOS One.

Responses:

Thank you for your suggestion. In response, we conducted an external validation of the paper's algorithm using the public dataset. The externally validated data were obtained from the public data platform, AIstudio, accessible via the Internet (https://aistudio.baidu.com/datasetdetail/273988). It is worth noting that the data used for external validation came from the public dataset and did not contain segmentation labels, so this external validation mainly evaluated the classification performance of ECRNet.

The dataset used for external validation consisted of 848 negative endometrial cytopathology images and 785 positive endometrial cytopathology images. The ratio of negative to positive images was 1.08:1. All images are papanicolaou stained images. During this external validation, we also chose four classification models (the ResNet-101 model, the DenseNet-121 model, the EfficientNet-B7 model, and the ResNeXt-101 model) to compare with our model. This is because these four models perform well in the classification task, second only to ECRNet. Specifically, these models above were first trained using images (1024 × 1024 pixels) from the XJTU-EC dataset instead of cell clumps (ROIs), and then tested directly on an external dataset. The specific results are shown in Table 6 as following:

TABLE 6. EXTERNAL VALIDATION COMPARISON RESULTS

Model Accuracy (%) Precision (%) Recall (%) F1-score (%)

ResNeXt-101 64.50 86.20 44.20 58.44

EfficientNet-B7 83.53 78.72 59.2 67.58

ResNet-101 73.50 100.00 53.10 69.37

DenseNet-121 80.10 77.20 67.20 71.90

Ours 95.32 94.57 96.17 95.37

In this external validation experiment, ResNet-101 has the highest precision of 100%, which means it has no false positives in the external validation. However, both ResNeXt-101, EfficientNet-B7 ResNet-101, and DenseNet-121 have low recall, which means they can easily miss screening positive patients. This is unacceptable for clinical tasks. In contrast, our model achieved the highest recall of 96.17%. In addition, ECRNet has the highest accuracy of 95.32%, followed by the DenseNet classifier with 83.53%. It is worth noting that ECRNet significantly outperforms the other four classifiers in terms of the F1 score. The F1 score is the reconciled average of precision and recall, which combines the information of precision and recall to provide a more comprehensive assessment of the performance of the classifiers, and the higher the F1 score, the better the performance of the classifiers. The higher the F1 score, the better the performance of the classifier. In summary, ECRNet has the best performance in this external validation experiment. The detailed changes (highlighted in red color) can be seen in Experiments and Results in the revised manuscript.

Finally, regarding the dataset of this manuscript, we are the first endometrial cytopathology dataset that contains segmentation and classification labels. Since making segmentation and classification labels is a laborious and time-consuming process, expanding the dataset will take more time, and we will definitely refine the information in the future. To the best of our knowledge, Koriakina et al. [1] proposed a deep multi-instance learning-based method for oral cancer detection, with data collected from a total of 24 patients; Guha et al. [2] collected 135 patients for a machine learning study of the liver; and Alhassan [3] proposed a deep learning-based method for breast cancer classification, trained and tested on a publicly available dataset of a total of 82 patients, with no external validation.

Previously, there was a lack of deep learning-based research on endometrial cancer cytopathology images, and our work is an initial exploration in automated endometrial cytopathology screening. In addition, our model can overcome the challenge of staining bias in clinical images and has been validated in external data with some significance.

Once again, we would like to express our gratitude for your understanding and sincerely hope for your approval.

References:

[1] Koriakina N, Sladoje N, Bašić V, et al. Deep multiple instance learning versus conventional deep single instance learning for interpretable oral cancer detection[J]. Plos one, 2024, 19(4): e0302169.

[2] Guha S, Ibrahim A, Wu Q, et al. Machine learning-based identification of contrast-enhancement phase of computed tomography scans[J]. Plos one, 2024, 19(2): e0294581.

[3] Alhassan A M. An improved breast cancer classification with hybrid chaotic sand cat and Remora Optimization feature selection algorithm[J]. Plos one, 2024, 19(4): e0300622.

Attachment

Submitted filename: Response to Reviewers.docx

pone.0306549.s004.docx (23.8KB, docx)

Decision Letter 3

Kazunori Nagasaka

11 Jun 2024

PONE-D-23-24291R3A Deep Learning Framework for Predicting Endometrial Cancer from Cytopathologic Images with Different Staining StylesPLOS ONE

Dear Dr. Wang,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Please submit your revised manuscript by Jul 22 2024 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

  • A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

  • A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

  • An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Kazunori Nagasaka

Academic Editor

PLOS ONE

Journal Requirements:

Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice.

Additional Editor Comments:

Dear Authors,

Thank you for your submission.

I have received very important suggestions from one of our reviewers.

Please consider their comments and submit your revised manuscript as far as you can.

Sincerely,

Kazunori Nagasaka

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #4: (No Response)

********** 

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #4: Partly

********** 

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #4: N/A

********** 

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #4: No

********** 

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #4: Yes

********** 

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #4: Thank you for your kind response to my repeated requests for revisions. I believe the results of the external validation are good enough to publish a paper on this AI and inquire about it with society. However, I would like you to add one more reference (reference URL) to the newly added sentence in line 438, which mentions "AIstudio, accessible via the Internet." Please include the URL along with the date of access, in the format of a literature citation. If this is completed, I recommend accepting this paper.

********** 

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #4: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

PLoS One. 2024 Jul 31;19(7):e0306549. doi: 10.1371/journal.pone.0306549.r008

Author response to Decision Letter 3


11 Jun 2024

Response to Reviewer #4:

Thank you for your kind response to my repeated requests for revisions. I believe the results of the external validation are good enough to publish a paper on this AI and inquire about it with society. However, I would like you to add one more reference (reference URL) to the newly added sentence in line 438, which mentions "AIstudio, accessible via the Internet." Please include the URL along with the date of access, in the format of a literature citation. If this is completed, I recommend accepting this paper.

Responses:

Thank you for your kind suggestion. In response, we have added the reference URL [69] to the newly added sentence in line 438, referencing the example format provided by PLOS One. Specific changes (highlighted in red) can be found in the revised manuscript and in the following annexes. Finally, we would like to thank you or providing us with valuable revision opportunities to improve the quality of our paper.

Experiments and Results:

In addition, we conducted an external validation of the paper's algorithm using the public dataset. The externally validated data were obtained from the public data platform, AIstudio, accessible via the Internet [69]. It is worth noting that the data used for external validation came from the public dataset and did not contain segmentation labels, so this external validation mainly evaluated the classification performance of ECRNet.

References:

69. Endometrial Cancer; 2024 [cited 2024 Jun 4]. Database: AIstudio [Internet]. Available from: https://aistudio.baidu.com/datasetdetail/273988.

Attachment

Submitted filename: Response to Reviewers.docx

pone.0306549.s005.docx (20.3KB, docx)

Decision Letter 4

Kazunori Nagasaka

19 Jun 2024

A Deep Learning Framework for Predicting Endometrial Cancer from Cytopathologic Images with Different Staining Styles

PONE-D-23-24291R4

Dear Dr. Wang,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice will be generated when your article is formally accepted. Please note, if your institution has a publishing partnership with PLOS and your article meets the relevant criteria, all or part of your publication costs will be covered. Please make sure your user information is up-to-date by logging into Editorial Manager at Editorial Manager® and clicking the ‘Update My Information' link at the top of the page. If you have any questions relating to publication charges, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Kazunori Nagasaka

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Dear Authors,

Thank you so much for your submission to Plos One.

I am pleased to tell you that your manuscirpt is acceptable for publication to our journal.

I think the manuscript is very significant and useful in the research area.

We look forward to your future manuscript.

Sincerely,

Kazunori Nagasaka

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #4: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #4: (No Response)

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #4: (No Response)

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #4: (No Response)

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #4: (No Response)

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #4: (No Response)

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #4: No

**********

Acceptance letter

Kazunori Nagasaka

25 Jun 2024

PONE-D-23-24291R4

PLOS ONE

Dear Dr. Wang,

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now being handed over to our production team.

At this stage, our production department will prepare your paper for publication. This includes ensuring the following:

* All references, tables, and figures are properly cited

* All relevant supporting information is included in the manuscript submission,

* There are no issues that prevent the paper from being properly typeset

If revisions are needed, the production department will contact you directly to resolve them. If no revisions are needed, you will receive an email when the publication date has been set. At this time, we do not offer pre-publication proofs to authors during production of the accepted work. Please keep in mind that we are working through a large volume of accepted articles, so please give us a few weeks to review your paper and let you know the next and final steps.

Lastly, if your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

If we can help with anything else, please email us at customercare@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Professor Kazunori Nagasaka

Academic Editor

PLOS ONE

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Supplementary Materials

    Attachment

    Submitted filename: review paper.docx

    pone.0306549.s001.docx (14.9KB, docx)
    Attachment

    Submitted filename: Response to Reviewers.docx

    pone.0306549.s002.docx (51.6KB, docx)
    Attachment

    Submitted filename: Response to Reviewers.docx

    pone.0306549.s003.docx (735.8KB, docx)
    Attachment

    Submitted filename: Response to Reviewers.docx

    pone.0306549.s004.docx (23.8KB, docx)
    Attachment

    Submitted filename: Response to Reviewers.docx

    pone.0306549.s005.docx (20.3KB, docx)

    Data Availability Statement

    Data cannot be shared publicly because of privacy protection. The owner of the data is the Ethics Committee of the First Affiliated Hospital of Xi’an Jiaotong University, and therefore it is not freely available. The Ethics Committee of the First Affiliated Hospital of Xi’an Jiaotong University imposed these restrictions. Data are available from the Ethics Committee (xjyfyllh@163.com) for researchers who meet the criteria for access to confidential data. The externally validated data were obtained from the public data platform, AIstudio, accessible via the Internet (https://aistudio.baidu.com/datasetdetail/273988).Anyone can access this data after registering an account on this platform.


    Articles from PLOS ONE are provided here courtesy of PLOS

    RESOURCES