Multitask deep learning models for ultrasound image analysis: identification of high-grade serous ovarian cancer and segmentation of tumor regions and intratumoral solid components

Abstract

Background

High-grade serous ovarian cancer (HGSOC) is characterized by high incidence and mortality rates, yet effective detection methods remain limited. In the present study, deep learning models capable of identifying HGSOC while concurrently providing segmentation results of tumor regions and intratumoral solid components were developed and validated.

Methods

This retrospective single-center diagnostic study included 395 patients (and their combined 1745 ultrasound images) with pathologically confirmed primary malignant ovarian tumors who underwent preoperative ultrasound examination between January 2018 and December 2022. The dataset was split into training (n = 237; 148 HGSOC, 89 Non-HGSOC), validation (n = 79; 49 HGSOC, 30 Non-HGSOC), and test (n = 79; 50 HGSOC, 29 Non-HGSOC) sets at a 6:2:2 ratio. Four models were constructed, all based on Residual Network-18 (ResNet-18) for feature extraction from grayscale and color Doppler ultrasound (US) images. Model A was built solely on image features, while Models B, C, and D further integrated additional features via a multilayer perceptron (MLP) as follows: Model B incorporated cancer antigen 125 (CA125); Model C combined CA125 and human epididymis protein 4 (HE4); and Model D integrated CA125, HE4, and age. All the models employed a dual-branch decoder for classification and segmentation to generate the corresponding outputs. The discriminative ability of the models was assessed using the area under the receiver operating characteristic curve (AUC), while classification performance was evaluated with the sensitivity and specificity. The Dice surface coefficient (DSC) was determined to evaluate segmentation performance.

Results

Among the 395 patients, 62.5% had HGSOC, while 37.5% were diagnosed with other pathological types. In the identification of HGSOC, the four models yielded AUC values of 0.72 (Model A), 0.81 (Model B), 0.80 (Model C), and 0.84 (Model D). Compared with Model A, which solely used image features, Model D resulted in the most significant increase in the AUC value (p = 0.03). The sensitivity across the models ranged from 0.66 to 0.86, with Model B yielding the highest value (0.86); the specificity values ranged from 0.63 to 0.93, with Model D yielding the highest value (0.93). The DSC values ranged from 0.70 to 0.86.

Conclusions

Deep learning models can be used to distinguish HGSOC and concurrently generate real-time segmentation results for relevant regions.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13048-026-02070-5.

Background

Ovarian cancer (OC) is among the most lethal gynecological carcinomas, with a five-year survival rate of less than 45% [1, 2], and thus seriously endangers women’s health. A prominent biological characteristic of OC is its marked intratumoral heterogeneity; notably, the prognosis and clinical management of OC are strongly associated with its pathological subtypes and disease stages [3, 4]. Accurate identification of the pathological type of OC is a core component of clinical diagnosis and personalized treatment. High-grade serous ovarian cancer (HGSOC) is the most prevalent histological subtype, accounting for approximately 70% of all OC cases [5]. Compared with other pathological subtypes, HGSOC exhibits more aggressive biological behaviors, including stronger invasiveness, higher susceptibility to drug resistance, and a greater likelihood of recurrence [6]. Furthermore, compared with those with other subtypes, patients with HGSOC exhibits differences in terms of neoadjuvant therapy eligibility, surgical radicality, targeted maintenance therapy, and prognostic monitoring [5, 7]. The accurate, preoperative identification of HGSOC would facilitate the optimization of treatment decisions and the precise formulation and implementation of treatment and follow-up protocols, thereby providing patients with personalized medical support and ultimately significantly improving their clinical outcomes.

Ultrasound (US) is the first choice of diagnostic modality for ovarian cancer [8]. Different histological subtypes exhibit distinct sonographic features, however: for example, endometrioid carcinoma primarily presents as large, unilateral masses and a low likelihood of ascites; clear cell carcinoma predominantly manifests as unilateral lesions, often accompanied by vascularized papillary projections and ground-glass echogenic cyst fluid; and HGSOC mainly presents as bilateral, advanced-stage lesions, predominantly exhibiting a solid or multilocular-solid morphology with rare papillary projections but a high likelihood of ascites [9–11]. Nevertheless, there is still some overlap between HGSOC and other pathological types, and so relying solely on subjective assessments to diagnose HGSOC is challenging, as there is currently no objective preoperative diagnostic criteria. In recent years, the rapid advancement of artificial intelligence (AI) technology has provided more objective and efficient approaches for the ultrasound diagnosis of ovarian cancer [12–14]. The findings of previous studies [15, 16] from our center have indicated that the use of deep neural networks (DNNs) can predict ovarian malignancy with a diagnostic accuracy comparable to that of expert human examiners. However, most existing studies focused on the classification of the benign or malignant nature of ovarian tumors, and only a few studies have explored pathological subtype diagnosis [17–19], let al.one the accurate diagnosis of HGSOC. In clinical practice, HGSOC is diagnosed on the basis of surgical pathological results, but this method is invasive [7]. To date, all efforts to implement an effective screening protocol for HGSOC have proven unsuccessful [5]. Consequently, the development of a method that can recognize HGSOC objectively and efficiently is urgently needed.

In this study, deep learning models capable of distinguishing HGSOC from other primary ovarian cancers while concurrently providing real-time segmentation results of tumor regions and the solid components within the tumor were trained and validated.

Methods

Study design

This was a retrospective, observational, single-center study conducted at the Department of Obstetrics and Gynecology at Ruijin Hospital in Shanghai, China. Patient inclusion criteria included surgical pathology of a primary malignant tumor of the ovary and transabdominal, transvaginal or transrectal ultrasound examination prior to the operation from January 2018 to December 2022. The exclusion criteria were poor US image quality, neoadjuvant chemotherapy prior to the ultrasound examination or a previous history of malignant ovarian tumors. In this study, we considered only primary malignant ovarian tumors in order to explore the internal differences among different pathological types. We did not include borderline tumors or metastatic tumors because they are relatively rare and have unique biological behaviors. In addition, most patients with metastatic tumors had undergone radiotherapy or chemotherapy for the primary disease before they underwent the US examination; this treatment may affect the internal features of the imaged lesions.

We divided the data into a training set, validation set, and test set at a ratio of 6:2:2. This study was approved by the Institutional Ethics Committee of Ruijin Hospital Affiliated to Shanghai Jiaotong University School of Medicine, with exemption from obtaining informed consent from individual patients (Grant No. 2025 − 347). The study protocol adheres to the core ethical requirements of the Declaration of Helsinki regarding medical research.

The medical records data used were obtained from the HIS system of our hospital and included patient age, menopausal status, pathological results, tumor staging, cancer antigen 125 (CA125) concentration, human epididymis protein 4 (HE4) concentration and imaging examination reports. The ultrasound images of the patients were downloaded from the PACS system. The images used for each patient included grayscale US images showing the plane with the maximal lesion dimension and its orthogonal plane, grayscale US images showing the maximum size of the solid component and its orthogonal plane (if there was a solid component) and color Doppler US images. The World Health Organization (WHO) classification [20] was used to define the tumor subtype, and the International Federation of Gynecology and Obstetrics criteria [21] was used to stage malignant tumors.

US examination

All US examinations were performed by using four US systems (GE Voluson E10, GE Healthcare; and Philips EPIQ Elite, EPIQ5 and A70, Philips Healthcare) with 5.0–9.0 MHz, 4.0–8.0 MHz, and 3.0–10.0 MHz transvaginal probes and 1.0–5.0 MHz transabdominal probes. If there were lesions in both ovaries or more than one lesion in one ovary, we evaluated the most complex or largest tumor, and only one tumor was evaluated per patient. The following sonographic features, determined on the basis of standardized terms, definitions, and measurements from the International Ovarian Tumor Analysis group [22] and the Ovarian-Adnexal Reporting and Data System (O-RADS) [23], were recorded: unilateral/bilateral, category (unilocular cyst, unilocular-solid cyst, multilocular cyst, multilocular-solid cyst, solid tumor), maximal diameter of the mass, maximal diameter of the largest solid component (height, 3 mm; arising from the cyst wall or septation and protrudes into the cyst cavity), presence or absence of papillary projections, presence or absence of ascites, presence of pelvic nodules, and color Doppler score (1, no flow; 2, minimal flow; 3, moderate flow; 4, very strong flow).

Image segmentation

Ultrasound images of the mass were saved in .jpg format and used in our analysis. The images were screened for quality by an expert gynecologist (H.C.). The region of interest (ROI) of each image was segmented manually by the author (J.S.) using an open-source labeling tool (LabelMe). The ROI was delineated to include the maximum diameter of the mass, maximum diameter of the solid portion and blood flow within the solid component of the mass.

Data processing

Before the images were fed into the deep learning model, the raw images underwent standardized preprocessing. Initially, image processing techniques were applied to extract the region within the ultrasound sector window contour, removing non-tissue artifacts such as scratches and caliper marks. The images were subsequently center-cropped to a uniform resolution of 224 × 224 pixels. To minimize the variability in physical resolution, organ position, and morphological differences during ultrasound acquisition, random data augmentation methods were applied, including horizontal or vertical translation (0–10%), rotation (0–15°), and horizontal flipping (50% probability). Additionally, all the images were normalized to have zero mean and unit variance to mitigate the impact of pixel value distribution differences across the images acquired by different ultrasound devices on model performance.

Model building

In this study, several multimodal, multitask deep learning models that integrate ultrasound grayscale images, color Doppler flow images, and clinical indicators were developed to simultaneously identify HGSOC and segment lesion regions as well as solid components within lesions (Fig. 1A). The model architecture comprises two major components: an encoder and a decoder. The encoding module is a feature extraction system composed of three sub encoders. Two image encoders extract features from grayscale US images and color Doppler US images, while the third sub encoder is a multilayer perceptron (MLP) [24], which serves as a clinical indicator encoder to extract clinical indicator features. To identify the most suitable image encoder, four networks—Residual Network-18 (ResNet-18) [25], ResNet-50, Densely Connected Convolutional Network (DenseNet) [26], and Vision Transformer (ViT) [27]—were pretrained exclusively on the two modalities of ultrasound imaging (Fig. 1B). After comprehensively evaluating their classification performance, segmentation effects, and computational efficiency on the ultrasound images, we ultimately selected ResNet-18 as the backbone image encoder (Table S1). The bimodal ultrasound imaging-based ResNet-18 was subsequently used as a baseline model, namely, Model A, for comparison with fusion models incorporating various clinical indicators. Specifically, we constructed three fusion models that sequentially added clinically significant indicators one by one, namely, Model B (based on the two modalities of ultrasound imaging and CA125 concentration), Model C (based on the two modalities of ultrasound imaging and CA125 and HE4 concentrations), and Model D (based on the two modalities of ultrasound imaging, CA125 and HE4 concentrations, and age). Features from each modality encoded by each sub encoder were concatenated and fused before being input into the decoding component, which. The decoding part includes two task branches: classification and segmentation. The classification branch maps the fused global features to a two-dimensional output space corresponding to the scores of the two types of labels and normalizes them into probability values through the softmax function to obtain the probability distribution of the disease malignancy for each patient. Finally, disease classification is performed for the case by comparison with a preset critical threshold determined with the goal of maximizing the Youden index to achieve the optimal balance between sensitivity and specificity. The segmentation branch was designed with a U-Net structure (Fig. 1C), where the up sampling stage of the decoder fuses features of different encoder layers through skip connections, enabling pixel-level segmentation of the lesion region and the solid component region.

Fig. 1.

Model structure diagram. A Network structure diagram of the multimodality models. B Structural diagram of the image encoders. C Structure of the segmentation branch

Training process

To ensure that the model parameters better adapt to the requirements of multitasking for disease classification and lesion segmentation, all the model weights were randomly initialized, adopting a training strategy from scratch. The initial learning rate was set at 0.001, the Adam optimizer was selected to adaptively adjust the learning rate, and a weight decay coefficient of 1e-4 was set to suppress overfitting. A two-stage optimization strategy was adopted to balance the classification and segmentation performance. The first stage involved training for 100 epochs to update all the model parameters so that the model converges on the segmentation task and achieves stable segmentation performance. In the second stage, the weights of the image encoder and the segmentation branch were frozen, and only the classification branch was fine-tuned until the loss converges and the classification performance stabilized without further improvement. During training, the cross-entropy loss function was used to optimize the classification task, with a label smoothing strategy introduced to mitigate overfitting to single-class labels by incorporating a fixed proportion of smoothing terms in the true label distribution, thereby enhancing model robustness and generalization. For the segmentation task, binary cross-entropy with logit loss (BCEWithLogits loss) was used to measure the pixel-level difference between the predicted mask and the true mask. Throughout the training process, both task losses were simultaneously optimized to achieve synergistic improvements in segmentation and classification performance.

Statistical analysis

All the data analyses were performed using IBM SPSS Statistics v29.0 and Python (version 3.8). Normality tests were conducted for continuous data. Chi-square tests, t tests, and Mann‒Whitney U tests were used to compare the ultrasonic characteristics of the tumors and the clinical characteristics of the patients. Categorical data are expressed as n (%), and continuous data are expressed as medians and interquartile ranges (25th and 75th percentiles). All the statistical calculations were performed with a 95% confidence interval (CI). To evaluate the classification performance of our developed deep learning models on the test set, we computed the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity values, the positive predictive value (PPV), the negative predictive value (NPV), the F1 score and the maximum Youden index. For segmentation performance assessment, we employed the Dice surface coefficient (DSC), where 0 indicates no overlap between the predicted and ground-truth contours and 1 represents a perfect surface match (ideal scenario). The closer the DSC value is to 1, the better the model’s segmentation performance. To determine whether the models that integrated clinical indicators outperformed the image feature-only baseline model (Model A), we utilized a bootstrapping approach to test the AUC metric (a detailed description of the bootstrapping approach is provided in the Supplementary Material). All tests were two-sided, with p values < 0.05 indicating statistical significance.

Visualization of the deep learning model

To better illustrate which part of the ultrasound image most impacts the classification result, we use the gradient-weighted class activation mapping technique (Grad-CAM) [28] to construct heatmaps depicting the regions of interest on which Model A (i.e., the model that uses ResNet-18 as the backbone for feature extraction and that was constructed solely on the basis of the two ultrasound imaging modalities) focused.

Results

From January 2018 to December 2022, 423 patients with one or more ovarian tumors were examined at the Department of Obstetrics and Gynecology at Ruijin Hospital in Shanghai, China. Among these patients, 28 were excluded because of poor US image quality (n = 9), the use of neoadjuvant chemotherapy prior to ultrasound examination (n = 13) or a previous history of malignant ovarian tumors (n = 6). Eventually, a total of 395 patients with 1,745 ultrasound images were included in the final analysis. A flowchart of the enrollment process and the procedure for model construction is shown in Fig. 2. The dataset was divided according to a 6:2:2 ratio, resulting in a training set (comprising 148 HGSOC and 89 other pathological types; totaling 1054 images; 60%), a validation set (consisting of 49 HGSOC and 30 other pathological types; comprising 342 images; 20%), and a test set (comprising 50 HGSOC and 29 other pathological types; encompassing 349 images; 20%).

Fig. 2.

Flowchart of enrollment and the procedure for model construction

Clinical, ultrasound and histopathological data

The pathological results of the patients are summarized in Table S2. Among these patients, 247 (62.5%) had HGSOC, while the remaining 148 (37.5%) had other pathological types of ovarian cancer. Clinical and ultrasound characteristics are presented in Table 1. There were no significant differences in demographic or clinical characteristics between the training, validation, and test sets, as detailed in Table 2. There were various differences between HGSOC and other pathological types. The mean age of patients with HGSOC was greater than that of patients with alternative pathological classifications, with median ages at diagnosis of 61 and 52 years, respectively (p < 0.001). Serum tumor marker levels were significantly higher in HGSOC patients than in patients with other pathological types, as reflected by the median CA125 levels (823.4 vs. 103.0, p < 0.001) and HE4 levels (287.0 vs. 64.0, p < 0.001). There are also significant differences in ultrasound characteristics between HGSOC and other ovarian cancers. HGSOC mostly manifested as a multilocular-solid cyst or solid tumor with bilateral occurrence. Papillary projections were less common. Instead, the incidence of ascites and pelvic nodules was notably high (p < 0.001).

Table 1.

Clinical and ultrasound characteristics of patients with and without HGSOC (n=395)

Characteristic	HGSOC(n=247)	Non-HGSOC(n=148)	P
Age (years)	61(52-67)	52(42-61)	<0.001
Menopausal Status	Pre/Post (56/191)	Pre/Post (67/81)	<0.001
CA125 (U/mL)	823.4(213.8-2438.8)	103(21.6-424.9)	<0.001
HE4(pmol/L)	287(118.5-579.9)	64(39-121)	<0.001
Tumor Stages			<0.001
I	24(9.7)	88(59.5)
II	18(7.3)	16(10.8)
III	113(45.7)	27(18.2)
IV	92(37.2)	17(11.5)
Unilateral/Bilateral	111/136	127/21	<0.001
Categories			<0.001
Unilocular cyst	0	0
Unilocular-solid cyst	16(6.5)	28(18.9)
Multilocular cyst	2(0.8)	1(0.7)
Multilocular-solid cyst	71(28.8)	74(50)
Solid tumor	158(64.0)	45(30.4)
Maximum lesion diameter (mm)	68(44-98)	107.5(70.5-157)	<0.001
Maximum largest solid component diameter (mm)	52(35.5-79)	57 (39-103)	0.037
Papillary projection	40(16.2)	47(31.8)	<0.001
Ascites	149(60.3)	41(27.7)	<0.001
Pelvic Nodules	115(46.6)	11(7.4)	<0.001
Color Doppler score			0.002
No flow, score 1	4(1.6)	5(3.4)
Minimal flow, score 2	26(10.5)	25(16.9)
Moderate flow, score 3	32(13.0)	29(19.6)
Very strong flow, score 4	185(74.9)	89(60.1)

Table 2.

Demographic and clinical characteristics of patients in the training set, validation set and test set (n = 395)

Characteristic	Training Set (n = 237)	Validation Set (n = 79)	Test Set (n = 79)	Total (n = 395)	P
HGSOC/non-HGSOC	148/89	49/30	50/29	247/148	0.986
Age (years)	58(50–66)	58(48–64)	56(48–68)	58(49–66)	1.215
Menopausal Status (Pre/Post)	74/163	27/52	22/57	123/272	0.691
CA125 (U/mL)	414.1 (71.6-1665.6)	299.5 (50.3-1541.2)	703.5 (133.5-2081.5)	428 (74-1715)	3.644
HE4(pmol/L)	123 (65–479)	123 (64–466)	159 (69–432)	123 (65.9–468)	0.378
Tumor Stages					0.799
I	67(28.3)	23(29.1)	22(27.9)	112(28.4)
II	25(10.6)	5(6.3)	4(5.1)	34(8.6)
III	81(34.2)	29(36.7)	30(38.0)	140(35.4)
IV	64(27.0)	22(27.9)	23(29.1)	109(27.6)
Unilateral/Bilateral	148/89	51/28	39/40	238/157	0.082
Categories					0.343
Unilocular cyst	0	0	0	0
Unilocular-solid cyst	22(9.3)	12(15.2)	10(12.7)	44(11.1)
Multilocular cyst	1(0.4)	2(2.5)	0	3(0.8)
Multilocular-solid cyst	90(38.0)	27(34.2)	28(35.4)	145(36.7)
Solid tumor	124(52.3)	38(48.1)	41(51.9)	203(51.4)
Maximum lesion diameter (mm)	77 (51-118.5)	84 (53–116)	83 (47–122)	80 (50–117)	0.332
Maximum largest solid component diameter (mm)	52 (37.3–82.8)	61 (38–88)	57 (35–86)	54 (38-85.5)	1.341
Papillary projection	55(23.2)	14(17.7)	18(22.8)	87(22.0)	0.585
Ascites	108(45.6)	39(49.4)	43(54.4)	190(48.1)	0.382
Pelvic Nodules	78(32.9)	21(26.6)	27(34.2)	126(31.9)	0.515
Color Doppler score					0.606
No flow, score 1	5(2.1)	2(2.5)	2(2.5)	9(2.3)
Minimal flow, score 2	34(14.4)	5(6.3)	12(15.2)	51(12.9)
Moderate flow, score 3	35(14.8)	12(15.2)	14(17.7)	61(15.4)
Very strong flow, score 4	163(68.8)	60(76.0)	51(64.6)	274(69.4)

The segmentation performance of the models

We tested the DSCs of different models on the test sets to verify their segmentation performance for lesions and solid components within lesions. As shown in Fig. 3, Model A achieved a DSC of 0.86 for lesion area segmentation and 0.77 for solid component area segmentation within lesions. After clinical indicator information was incorporated, the model’s segmentation performance slightly decreased. Model B achieved a DSC of 0.85 for lesion areas and 0.76 for solid component areas. Model C, which built upon Model B with the HE4 concentration, exhibited DSCs of 0.83 and 0.73 for the two target areas, respectively. When age was further included, the DSCs of Model D for the two target areas were 0.81 and 0.70. Relatively speaking, each model can serve as an effective tool to guide physicians in focusing on lesion areas and solid components within lesions. Additionally, they can serve as alternatives to manual segmentation of ROIs. Examples of the segmentation results of Model A for the two target areas are shown in Fig. 4.

Fig. 3.

Segmentation performance of different models in the testing set

Fig. 4.

Examples of the segmentation results of Model A for the two target areas

Pathology type classification performance of the models

Table 3 compares the efficacy of different models in identifying HGSOC in the test set. Moreover, their performance across all three datasets (the training set, validation set, and test set) is shown in Table S3. The evaluation metrics used included the area under the curve (AUC), sensitivity, and specificity values, the positive predictive value (PPV), the negative predictive value (NPV), the F1 score, the maximum Youden index and the cutoff value. The comparison of AUC curves for different models is shown in Fig. 5. The x-axis represents the false positive rate (FPR), and the y-axis represents the true positive rate (TPR). Model A achieved an AUC of 0.72. After the clinical indicators were sequentially incorporated, compared with those of Model A, the AUC values of the three models (Model B, Model C, and Model D) increased to 0.81, 0.80, and 0.84, respectively. Among them, Model D, which incorporated the most clinical indicators, achieved the most significant improvement in classification performance (AUC: 0.84 vs. 0.72, p = 0.03). In terms of sensitivity, Model B, which incorporates only CA125 concentration, had the highest sensitivity, reaching 0.86, followed by Model C at 0.84 and Model D at 0.70, with Model A having the lowest sensitivity. Model D, which incorporates the most clinical indicators, has the highest specificity, reaching 0.93, followed by Model A, Model B, and Model C in that order. In terms of accuracy, the values for Model B, Model C, and Model D were 0.79, 0.76, and 0.78, respectively, all exceeding that of Model A (0.69). The maximum Youden index, a measure of overall diagnostic performance, was highest for Model D at 0.65. In addition, the F1 scores of the respective models range from 0.73 to 0.84. Cutoff values were determined for each model, with values ranging from > 0.48 to > 0.65.

Table 3.

Comparison of the efficacy of different models in the testing set

Model	AUC	Accuracy	Sensitivity	Specificity	NPV	PPV	F1 score	Maximum Youden index	Cutoff value	p
Model A	0.72(0.58–0.83)	0.69(0.57–0.79)	0.66(0.51–0.79)	0.74(0.57–0.89)	0.56(0.42–0.70)	0.81(0.68–0.94)	0.73(0.62–0.83)	0.44	>0.51	NA
Model B	0.81(0.69–0.91)	0.79(0.70–0.87)	0.86(0.76–0.94)	0.67(0.47–0.84)	0.74(0.54–0.94)	0.82(0.71–0.93)	0.84(0.75–0.91)	0.60	>0.65	0.178*
Model C	0.80(0.69–0.91)	0.76(0.67–0.86)	0.84(0.73–0.94)	0.63(0.44–0.81)	0.70(0.50–0.89)	0.80(0.67–0.92)	0.82(0.73–0.89)	0.53	>0.58	0.128*
Model D	0.84(0.74–0.92)	0.78(0.69–0.87)	0.70(0.57–0.83)	0.93(0.82–1.00)	0.64(0.46–0.82)	0.95(0.87–0.99)	0.81(0.71–0.89)	0.65	>0.48	0.03*

Fig. 5.

Receiver operating characteristic (ROC) curves summarizing the performance in the identification of HGSOC. Model A: based only on the two ultrasound imaging modalities. Model B: based on the two ultrasound imaging modalities + CA125 concentration. Model C: based on the two ultrasound imaging modalities + CA125 concentration + HE4 concentration. Model D: based on the two ultrasound imaging modalities + CA125 concentration + HE4 concentration + age

As shown in Fig. 6A, Model A failed to accurately distinguish HGSOC solely on the basis of image features, whereas the subsequent models incorporating tumor markers achieved effective classification. Nevertheless, Model D still exhibited misclassification. As shown in Fig. 6B, a 15-year-old non-HGSOC patient was correctly classified by Model A but misclassified by Model B and Model C because of elevated tumor marker levels. However, given that the patient was only 15 years old, Model D accurately reclassified the patient’s disease as a non-HGSOC. A 48-year-old HGSOC patient with elevated tumor marker levels is shown in Fig. 6C; for this patient, Models A, B, and C correctly identified the patient’s disease, whereas Model D ultimately misclassified it as non-HGSOC.

Fig. 6.

Examples of three patients.A A 50-year-old HGSOC patient with CA125 and HE4 concentrations of 2499.2 U/mL and 1468.9 pmol/L, respectively. Model A failed to classify the disease correctly, whereas Models B and C successfully identified the disease as HGSOC, while Model D misclassified it as non-HGSOC. B A 15-year-old non-HGSOC patient with CA125 and HE4 concentrations of 524.9 U/mL and 48 pmol/L, respectively. Model A correctly classified the disease as non-HGSOC, whereas Models B and C misclassified the disease as HGSOC, and Model D corrected the classification to non-HGSOC. C A 48-year-old HGSOC patient with CA125 and HE4 concentrations of 1001.6 U/mL and 95 pmol/L, respectively. Models A, B and C all successfully classified the disease as HGSOC, whereas Model D classified it as non-HGSOC

Calibration curves of the models

The calibration curves of the four models evaluated in the test set are shown in Fig. 7. For Model A, pronounced risk underestimation is observed in the low predicted probability regime; nevertheless, the curve converges closely to the ideal calibration line in the high-probability regime, yielding strong overall calibration performance. Model B displays marked deviation: within the low-to-moderate predicted probability interval, it drastically overestimates the event probability, with the actual positive fraction converging to the predicted value only in the high predicted probability interval. Model C similarly shows overestimation in the low predicted probability interval. Additionally, periodic plateau regions emerge in its curve, leading to slight fluctuations in the calibration stability. For Model D, the event occurrence probability is underestimated in the low predicted probability interval. As the predicted probability increases in the high-probability regime, the curve gradually approximates the perfect calibration line.

Fig. 7.

Calibration curves of the models

Channel attention visualization analysis

As illustrated in Fig. 8, we used Grad-CAM to generate heatmaps to depict the regions focused on by the model during decision-making. In the Grad-CAM images, the activated (reddish bias) regions are heavily weighted during the prediction of the final results, whereas the blue regions are generally not factored into the outcome. The results demonstrated that the model exhibited a greater focus on the solid components of the tumor, from which it extracted relevant features for classification.

Fig. 8.

Ultrasound images and heatmaps of the two cases. A A 69-year-old female with a 120 mm multilocular-solid cyst mass in the left adnexal region. The pathological diagnosis was HGSOC. B 74-year-old female with a 38 mm multilocular-solid cyst mass in the right adnexal region. The pathological diagnosis was clear cell carcinoma

As shown in Fig. 8A, there was a 69-year-old female with a 120 mm multilocular-solid cyst mass in the left adnexal region. The solid component of the mass shows hypo echogenicity, while the cystic component presents as anechoic areas with echogenic septa, forming multiple locules. No obvious papillary projections were observed, and the solid region exhibited abundant blood flow signals. Through image analysis, Model A revealed that the patient had a 67.1% probability of having HGSOC, which was confirmed in the final pathological assessment of the patient.

The case of a 74-year-old female patient is shown in Fig. 8B. A 38 mm multilocular-solid cyst mass was visualized in the right adnexal region, with the maximum diameter of the internal solid component measuring 13 mm, and moderate blood flow signals were detected within the solid component. According to Model A, the probability of the patient having HGSOC was 63%. When this exceeded the diagnostic threshold of 0.51, the patient was classified as having HGSOC. Tumor marker tests for the patient revealed a CA125 concentration of 2.4 U/ml and an HE4 concentration of 69 pmol/L. All 3 subsequent models that incorporated clinical indicators classified the disease as non-HGSOC. Ultimately, the patient’s pathological diagnosis was clear cell carcinoma.

Discussion

Clinical background

HGSOC is the most prevalent and deadly form of OC, defined by its aggressive biological behavior, high recurrence rate, and status as a primary contributor to the high mortality burden associated with OC [5]. Clinically, US is the preferred initial imaging modality, providing critical preliminary evidence for the suspicion of HGSOC [8]. In the context of serological detection, CA125 and HE4 have been established as core serum tumor biomarkers for HGSOC and play multifaceted and valuable roles in the diagnostic workflow—both individually and in combination [29]. In addition to imaging and serological indicators, age is a prominent and well-documented risk factor for HGSOC. The incidence of this malignancy increases significantly with advancing age, particularly in postmenopausal women [30].

Comparative analysis of the four models

In this study, we constructed four models for distinguishing HGSOC from other primary ovarian cancers. These four models exhibited distinct characteristics with respective focuses on the basis of different variable combinations, and the selection of models should be tailored to the core requirements of specific clinical scenarios. As a baseline model utilizing only two modalities of ultrasound images, Model A demonstrated the poorest overall performance, thus serving merely as a temporary preliminary screening reference when no biomarker data are available. Model B, which incorporates CA125 concentration in addition to the two modalities of ultrasound images, achieved the highest sensitivity (SEN = 0.86) among the four models and is suitable for preliminary screening scenarios in which the reduction of missed diagnoses is the priority. However, this model suffers from low specificity and large fluctuations in calibration, necessitating subsequent review procedures to mitigate misdiagnosis. Model C, with the further addition of HE4 concentration, exhibited relatively balanced metrics across all the performance indicators. Consequently, it is well suited for preliminary clinical decision-making scenarios. Model D, which integrates age as an additional variable, had the best discriminative ability (AUC = 0.84) and specificity (SPE = 0.93), with a statistically significant increase in the AUC value compared with that of Model A (0.84 vs. 0.72, p = 0.03). It is therefore applicable in precise diagnostic scenarios where the goal is to minimize unnecessary interventions, although it also results in a calibration bias of risk underestimation, requiring an upward adjustment of predicted values during clinical application.

We evaluated four models across representative cases (as illustrated in Fig. 6). Although Model D achieved the highest specificity, its sensitivity was low. We hypothesize that this phenomenon originates from the dual effects of incorporating clinical features. On the one hand, these features overcome inherent information gaps in ultrasound images and improve predictive accuracy via multimodality fusion. However, they may induce overfitting driven by feature redundancy and multicollinearity. The incidence of certain other pathological subtypes of ovarian cancer is relatively high among younger patients [31, 32]. Among the patients enrolled in this study, the median age at onset of non-HGSOC patients was significantly lower than that of HGSOC patients (52 vs. 61, p < 0.001). Given this characteristic, the model may acquire the associative pattern that “younger age corresponds to the non-HGSOC subtype”, which may introduce bias into the model’s learning process. However, a subset of younger HGSOC patients was also present in this study sample. Thus, incorporating the age indicator significantly improved the model’s specificity but may increase the number of missed diagnoses of special cases (e.g., young HGSOC patients), resulting in reduced sensitivity.

Strengths of the study

One advantage of this study is the novel application of DNNs to extract more in-depth features for HGSOC diagnosis. Several studies have explored the application of deep learning in distinguishing OC subtypes [14, 18, 33]. For instance, Wang et al. [18] evaluated the efficacy of DNNs in identifying benign, borderline, and malignant serous tumors using ultrasound images. Another international multicenter study [14] that validated the robustness of the Transformer-based model further confirmed its potential in distinguishing specific histological subtypes. Notably, however, none of the aforementioned studies conducted dedicated validation for the diagnosis of HGSOC. Prior pilot studies have utilized handcrafted image descriptors combined with machine learning classifiers for related diagnosis tasks [17, 34, 35]. For instance, Yang et al. [17] extracted radiomics features from manually segmented images and integrated clinical markers (e.g., CA125, HE4, and CA19-9) to develop models distinguishing HGSOC and non-HGSOC. The results revealed that combining recursive feature elimination (RFE) and a gradient boosting decision tree (GBDT) performed best, with an AUC of 0.87 and high sensitivity (0.88), specificity (0.75), and accuracy (0.82). Two other studies [34, 35] adopted similar approaches to construct models distinguishing between type I (non-HGSOC) and type II (predominantly HGSOC) subtypes of epithelial ovarian cancer. The first [34] reported a combined model with an AUC value of 0.82 in the test set (0.75 sensitivity, 0.85 specificity); the second [35] achieved an AUC value of 0.89, with calibration and decision curves demonstrating good consistency and clinical utility. While promising, these studies all share the key limitation of a reliance on manual tumor segmentation. By contrast, our models require physicians to manually outline the ROI during training, but for validation and test set inference and subsequent real-world application, it can directly output results from raw images. Furthermore, we employed Grad-CAM to visualize the image regions that the model attends to during decision-making. The results demonstrated that the model placed greater emphasis on the solid components of the tumor, from which relevant features were extracted for classification.

Another advantage is that our models can perform multiple tasks: while classifying pathological types, they can also provide real-time segmentation results of tumor regions and solid components within the tumor. Intratumoral solid components are regarded as key regions reflecting tumor biological characteristics, as they may harbor imaging features highly relevant to lesion nature [12]. All the models achieved DSCs ranging from 0.70 to 0.86. These findings confirm that our models can efficiently guide physicians to focus on objective regions and may serve as alternative tools to replace manual segmentation. However, the model’s segmentation performance decreased after the clinical indicators were incorporated. We speculate that the reason for this phenomenon may be that clinical indicators have a stronger correlation with the classification task, leading the multitask model to prioritize classification over segmentation, and their inherent noise and heterogeneity further hinder the model’s learning of key segmentation-essential spatial features [36, 37]. Notably, unlike other multitask models that use cascaded architectures to sequentially link segmentation and classification modules [13, 38], our method integrates multimodal features into separate classification and segmentation branches, enabling simultaneous dual-task output. Clinicians can simultaneously determine lesion anatomical localization and pathological stratification, optimize diagnostic workflows, reduce decision-making time, and enable timely intervention for suspected HGSOC patients. Furthermore, this model design approach can synergistically enhance the performance of both the classification branch and the segmentation branch: local spatial features from the segmentation branch enhance the recognition of the key region of the classification branch via spatial semantic input, whereas global semantic information from the classification branch refines the lesion boundary detection of the segmentation branch [39]. Additionally, this end-to-end training also avoids cumulative segmentation error transfer to classification in stepwise processing and reduces computational overhead from redundant feature extraction [40].

Limitations of the study

Nevertheless, our study has several limitations. First, this was a single-center, retrospective study and lacked validation through multicenter, prospective investigations. This design inherently entails biases from retrospective data collection and limited sample representativeness. Second, the survival rate of HGSOC patients is closely related to disease stage [4]. Owing to the limited sample size, we were unable to further conduct a diagnosis of early-stage HGSOC in this study. In the future, we will collect more samples to explore this possibility.

Overall, our study demonstrates the potential of DNNs to accurately identify HGSOC and provide real-time segmentation results.

Conclusions

In conclusion, the models trained and internally validated in this study are capable of accurately identifying HGSOC among malignant ovarian tumors and generating real-time segmentation results. These findings are expected to serve as valuable references for the achievement of pathological-level accuracy in deep neural network-assisted ultrasound diagnosis and the development of related multitask models.

Abbreviations

AI

Artificial Intelligence

AUC

Area Under the Curve

CA125

Cancer Antigen 125

CI

Confidence Interval

DNNs

Deep neural networks

DenseNet

Densely Connected Convolutional Network

DSC

Dice surface coefficient

Grad-CAM

Gradient-weighted Class Activation Mapping Technique

GBDT

Gradient Boosting Decision Tree

FPR

False Positive Rate

HE4

Human Epididymis Protein 4

HGSOC

High-grade Serous Ovarian Cancer

IOTA

International Ovarian Tumor Analysis

MLP

Multi-Layer Perceptron

OC

Ovarian Cancer

O-RADS

Ovarian-Adnexal Reporting and Data System

PPV

Positive Predictive Value

ResNet

Residual Network

RFE

Recursive Feature Elimination

ROI

Region of Interest

TPR

True Positive Rate

TVUS

Transvaginal Ultrasound

US

Ultrasound

ViT

Vision Transformer

WHO

World Health Organization

Authors’ contributions

CH and JC conceptualized and designed the study, supervised data collection, model training and validation, as well as reviewed and revised the manuscript. SJ and DJY collected data, carried out the initial analyses, drafted the initial manuscript, and revised the manuscript. FYH and JRS processed the data, utilized deep learning models, and revised the manuscript for deep learning models’ content. NYN and HL acquired the ultrasound images. All authors approved the final manuscript as submitted and agree to be accountable for all aspects of the work.

Funding

This study was supported by the Medical Innovation Project of the Shanghai Science and Technology Commission(20Y11914000).

Data availability

The images used in this study are available from the corresponding author CH upon request. All data analyzed in this study are included in the published article.

Declarations

Ethics approval and consent to participate

The study was approved by the Institutional Ethics Committee of Ruijin Hospital Affiliated to Shanghai Jiaotong University School of Medicine, with exemption from obtaining informed consent from individual patients (Grant No. 2025 − 347). The study protocol adheres to the core ethical requirements of the Declaration of Helsinki regarding medical research.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.