Abstract
Background/Aims
Lymphadenopathy presents diagnostic challenges, particularly for the mediastinal and intra-abdominal lymph nodes (LNs). Endoscopic ultrasonography (EUS) has emerged as a tool for LN detection; however, its accuracy varies. To enhance the diagnostic performance and minimize medical costs, assisting LN assessment using EUS is necessary. Machine learning (ML) offers potential for medical image analysis. This study aimed to develop an ML model for classifying mediastinal and intra-abdominal LNs using gastrointestinal EUS.
Methods
EUS images of mediastinal and intra-abdominal LNs were randomly split into training and validation datasets. U-Net was selected for LN segmentation, and six deep-learning architectures were combined with the k-nearest-neighbor algorithm for LN classification. Physicians, comprising one expert group and one trainee group, reviewed the validation dataset and made individual diagnoses. A logistic regression model was generated based on LN features. We compared the diagnostic yields of ML, expert and trainee groups, logistic regression analysis, and a combination of the various methods mentioned above for diagnosing LNs.
Results
In total, 93 patients were enrolled, providing 630 images. The ResNet-50+logistic regression analysis+expert group achieved the best F1 score and sensitivity of 0.89 and 100.0%, respectively. Paired comparisons revealed that the combination outperformed both experts and trainees in terms of the area under the curve (p<0.01).
Conclusions
ML assists in predicting the mediastinal and intra-abdominal LNs based on gastrointestinal EUS images, particularly when combined with expert expertise and logistic regression models.
Keywords: Deep learning, Endosonography, Lymphadenopathy
Graphical abstract
INTRODUCTION
Lymphadenopathy can result from hyperplasia, lymphoma, or tumor metastasis and has different prognoses and management strategies.1,2 Unlike superficial lymph nodes (LNs), diagnosing mediastinal and intra-abdominal LNs is challenging. Endoscopic ultrasonography (EUS) is a new detection method. Catalano et al.3 investigated the ability of EUS to predict mediastinal LNs in patients with esophageal cancer and identified hypoechogenicity, a sharp border, a round shape, and a size >10 mm as significant characteristics of malignancy. However, 20% of the patients lacking the aforementioned features were confirmed to have LN metastasis. Similarly, Gill et al.4 have reported that a round shape, sharp border, and short axis >8.3 mm were significant predictive factors of malignant mediastinal LNs in patients with lung cancer. When all these characteristics were present, the probability of malignancy was approximately 63%. Collectively, the diagnostic accuracy of EUS for mediastinal and intra-abdominal LNs ranges from 50% to 100% in various diseases, such as esophageal, gastric, pancreatic, and rectal cancers, demonstrating significant variability.5,6 In summary, the diagnosis of LNs using EUS, which is affected by the observer and the equipment, has inherent limitations. EUS-guided fine-needle aspiration/biopsy (EUS-FNA/B) confers the advantages of pathological elucidation and minimal invasiveness, although it requires the selection of the target LN for pathological examination.7 To reduce complications and medical costs, a tool that aids in predicting LNs using EUS needs to be identified.
Machine learning (ML) has gained increasing attention in medical image processing. It learns features from labeled data using artificial neural networks and interprets undefined data. Previous studies have combined ML with endobronchial ultrasound to identify intrathoracic LNs and have demonstrated promising results.8,9 Nevertheless, to the best of our knowledge, no ML model for classifying mediastinal and intra-abdominal LNs using gastrointestinal EUS has been developed.10-12 Therefore, this trial aimed to explore the value of ML in the prediction of lymphadenopathy based on gastrointestinal EUS images.
METHODS
Patients and datasets
The eligibility criteria were as follows: (1) patients who underwent EUS to diagnose mediastinal and intra-abdominal lymphadenopathies discovered during ultrasonography (US) or computed tomography (CT) between September 2016 and January 2023; (2) patients with EUS B-mode images without color/power Doppler; and (3) patients who provided informed consent. The exclusion criteria were as follows: (1) poor image quality, (2) history of radiotherapy, chemotherapy, or targeted therapy, and (3) failure to establish a definite diagnosis. Poor-quality images met the following conditions: (1) images containing comments or size measurements, and (2) images of LNs after EUS-FNA/B. To maximize the sample size, images from each lesion were not restricted. All images were randomly categorized into training and validation datasets. To mitigate heterogeneity among the images, preprocessing was conducted by ML experts, which involved fine-tuning the parameters and employing noise reduction filters.
EUS procedure
EUS procedures were performed by an endoscopist with at least 1,000 cases of experience using a scope (EG-580UT; Fujifilm Corporation), ultrasound system (SU9000; Fujifilm Corporation), and needles (Cook Endoscopy). Prior to the procedure, anesthesia was administered to the throat, and an intramuscular injection of 5 to 10 mg of diazepam was administered. The endoscopists observed the morphology, border, echo characteristics, and growth of the LN using a 7.5 to 12 MHz ultrasound probe. Subsequently, blood flow around the LN and surrounding tissues was assessed using Doppler imaging. If necessary, a puncture was made for pathological examination. Throughout the procedure, endoscopists captured images and recorded videos, if available.
Final diagnosis
LNs meeting at least one of the following conditions were considered benign: (1) pathology based on EUS-FNA/B or surgical procedure demonstrated no malignancy, and US/CT indicated no advancement without intervention after a 6-month follow-up (Fig. 1)8; and (2) absence of pathology, and US/CT indicated no advancement without intervention after a 6-month follow-up. LNs were considered malignant if they were confirmed based on pathology (Fig. 2) or in the case of benign pathological results, whereas US/CT indicated that the short axis of at least one LN had increased by >30% after a 6-month follow-up.13
Fig. 1.
Hematoxylin and eosin staining of the mesenteric lymph node in benign hyperplasia (×40).
Fig. 2.
Hematoxylin and eosin staining of lymph nodes with adenocarcinoma metastasis (×40).
LN segmentation model
Open-source Labelme software (https://github.com/wkentaro/labelme) was employed by an endoscopist (with at least 1,000 cases of experience in EUS operations) to create regions of interest in the EUS images. The labeled samples were used to train a U-Net model, a U-shaped encoder–decoder architecture that employs progressive downsampling to reduce image resolution and extract high-level features, and upsampling to restore image resolution for precise localization.14 It uses skip connections to integrate local image details (e.g., edges and textures) from early layers with hierarchical features (e.g., organ structures) from later layers, enabling automated segmentation of LNs in EUS images (Fig. 3). The model was trained using the Adam optimizer with an initial learning rate of 0.001 for 60 epochs, employing a batch size of 1 and resizing each image to 256×256 pixels. During training, U-Net employs a Dice loss function to optimize the segmentation performance, which is particularly effective for handling class imbalance.
Fig. 3.
Lymph node segmentation based on U-Net algorithm.
Development of the classifiers
We explored six deep-learning (DL) architectures: ResNet-50, ResNet-101, ResNet-152, DenseNet-121, DenseNet-169, and DenseNet-201. They rely on backpropagation (an algorithm that distributes error signals from the output layer backward through the network using chain rule differentiation) for gradient computation and gradient descent (an optimization method that minimizes the objective function by computing the gradients of the loss function with respect to the weights and subsequently updating the weights in the negative gradient direction) for parameter updates. In this process, the backpropagated gradients can experience exponential decay across multiple layers, a phenomenon known as the vanishing gradient problem, which severely impedes efficiency. Similar to U-Net, ResNets utilize residual blocks with shortcut connections that skip one or more layers, mitigate the vanishing gradient problem, and improve training stability. DenseNet alleviates information loss through dense connections, where each layer directly receives and concatenates feature maps from all preceding layers. Unlike DL, the k-nearest neighbor (k-NN) is a classical ML algorithm. It identifies the k-NNs of a query instance in the training dataset and predicts its class based on the majority vote or weighted average of the labels of these neighbors. Six DL architectures were individually combined with the k-NN algorithm, yielding six advanced classifiers.15 Finally, the LNs in the validation dataset were depicted using U-Net, which was used to test the efficacy of the six classifiers. Clinical information and final diagnoses were not available for the ML performers.
Human diagnosis
The physicians were divided into an expert group (two physicians with at least 1000 cases of experience in EUS operations) and a trainee group (two physicians with less than 100 cases of experience in EUS operations). Before the investigation, the trainees underwent a 3-month practice conducted by experts, which focused on the EUS operation and a literature review on the ultrasonographic characteristics of LNs. All physicians independently diagnosed LNs in the validation dataset and were blinded to the final diagnoses. They synthesized the clinical presentation, medical history, laboratory findings, and imaging characteristics to form individual diagnoses. To mitigate overestimation of the difference in efficacy between ML and humans, a correct diagnosis was considered for analysis if at least one member made an accurate diagnosis within the same group.
Statistical analysis
A logistic regression model was developed for LN features in the EUS images, which were independently extracted by two experts (with at least 1000 cases of experience in EUS operations). To reduce bias, each lesion was collected only once. Disagreements were resolved by another author (with at least 1,200 cases of experience in EUS). We processed the variables listed in Table 1. Variables with a p<0.10 after univariate analysis were retained for further analysis, with a threshold of 0.5 for discrimination.
Table 1.
Values of variables for logistic regression analysis
| Variable | Value |
|
|---|---|---|
| 0 | 1 | |
| Long axis (mm) | <20 | ≥20 |
| Short axis (mm) | <10 | ≥10 |
| S/L ratio | ≤0.5 | >0.5 |
| Shape | Irregular | Regular |
| Margin | Poorly demarcated | Sharply demarcated |
| Hilum | Visible | Absent |
| Texture | Homogeneous | Heterogeneous |
| Growth | Solitary or scattered | Conglomerations or fusions |
| Echogenicity | Hyperechoic | Hypoechoic |
S/L ratio, the long-to-short axis ratio.
Based on the validation dataset, we compared the diagnostic efficacy of ML, human experience, and logistic regression models as well as the performance of various combined approaches. We employed a strategy where the test result was deemed positive if any of the diagnostic methods indicated a positive outcome. The primary outcome was the F1 score, and the secondary outcomes focused on the area under the receiver operating characteristic curve (AUC). In addition, we assessed the sensitivity, specificity, accuracy, and positive and negative predictive values. All statistical analyses were performed using the IBM SPSS Statistics ver. 26.0 (IBM Corp.).
Ethical approval
This single-center retrospective study was conducted in accordance with the principles of the Declaration of Helsinki and was approved by the Ethics Committee of the Second Affiliated Hospital of Soochow University (JD-HG-2023-45).
RESULTS
Clinical characteristics
Initially, 122 patients with 838 EUS images of the LNs met the inclusion criteria. After eliminating cases with poor image quality and loss to follow-up, 630 images were randomly divided into a training dataset (200 benign LNs and 325 malignant LNs) and a validation dataset (40 benign LNs and 65 malignant LNs). The flow of participants and study design are depicted in Figure 4. Characteristics of the enrolled patients and LNs are summarized in Tables 2 and 3, respectively.
Fig. 4.
Diagram of patient selection and study design.
Table 2.
Characteristics of patients
| Characteristic | Value (n=93) |
|---|---|
| Age (median [IQR], yr) | 62.00 (52.7, 68.2) |
| Sex (male/female, n) | 51/42 |
| Diagnosis | |
| Benign | 29 |
| Inflammation | 15 |
| Reactive hyperplasia | 14 |
| Leiomyoma | 1 |
| Lipoma | 1 |
| Castleman's disease | 1 |
| Primary biliary cholangitis | 1 |
| Others | 10 |
| Malignant | 64 |
| Lymphoma | 19 |
| Pancreatic adenocarcinoma | 12 |
| Lung cancer | 8 |
| Gastric adenocarcinoma | 8 |
| Gallbladder carcinoma | 4 |
| Hepatocellular carcinoma | 3 |
| Cholangiocarcinoma | 3 |
| Pancreatic neuroendocrine tumor | 2 |
| Esophageal carcinoma | 2 |
| Colorectal carcinoma | 1 |
| Breast cancer | 1 |
| Ovarian cancer | 1 |
IQR, inter-quartile range.
Table 3.
Characteristics of lymph nodes
| Characteristic | Malignant (n=121) | Benign (n=79) |
|---|---|---|
| Region | ||
| Mediastinum | 20 | 26 |
| Abdomen | 101 | 53 |
| Long axis (median [IQR], mm) | 16.00 (14.5, 21) | 12.00 (9, 15) |
| Short axis (median [IQR], mm) | 10.00 (8.5, 14) | 8.00 (6, 10) |
| S/L ratio (median [IQR]) | 0.75 (0.63, 0.83) | 0.75 (0.63, 0.83) |
| Shape | ||
| Irregular | 62 | 32 |
| Regular | 59 | 47 |
| Margin | ||
| Poorly demarcated | 42 | 44 |
| Sharply demarcated | 79 | 35 |
| Hilum | ||
| Visible | 29 | 50 |
| Invisible | 92 | 29 |
| Texture | ||
| Homogeneous | 46 | 56 |
| Heterogeneous | 75 | 23 |
| Growth | ||
| Solitary or scattered | 40 | 46 |
| Conglomerations or fusions | 81 | 33 |
| Echogenicity | ||
| Hyperechoic | 56 | 44 |
| Hypoechoic | 65 | 35 |
IQR, inter-quartile range; S/L ratio, the long-to-short axis ratio.
Performance of the ML, human, and logistic regression models
Segmentation by U-Net was applicable for further classification, despite several errors (Fig. 5). All classifiers achieved an accuracy of >75.0% (Table 4). ResNet-50 had the highest F1 score (0.89) among the six classifiers (Table 5). The kappa value was 0.683, indicating substantial agreement with the final diagnosis. The average inference time for dealing with a single image was 0.66s.
Fig. 5.
Areas on the leftmost represent the markers by experts and areas on the rightmost represent the segmentation results by U-Net. (A) The lymph nodes are depicted accurately. (B, C) U-Net failed to identify the hilum and lesion with surrounding interference.
Table 4.
Best performance of six deep-learning algorithms
| Classifier | Best accuracy (%) | Best distance metric | Best k value |
|---|---|---|---|
| ResNet-50 | 85.7 | Canberra | 7 |
| DenseNet-201 | 83.8 | Euclidean | 7 |
| DenseNet-169 | 80.0 | Cosine | 7 |
| DenseNet-121 | 77.1 | Canberra | 7 |
| ResNet-101 | 76.2 | Euclidean | 1 |
| ResNet-152 | 75.2 | Euclidean | 1 |
Table 5.
ResNet-50 and final diagnosis of lymph nodes
| ResNet-50 | Final diagnosis |
Total | |
|---|---|---|---|
| Malignant | Benign | ||
| Malignant | 62 | 12 | 74 |
| Benign | 3 | 28 | 31 |
| Total | 65 | 40 | 105 |
The expert group recognized 48 (two lesions were misdiagnosed by ResNet-50) of the 65 malignant cases and 26 of the 40 benign cases, whereas the trainee group recognized 38 of the 65 malignant cases and 22 of the 40 benign cases. The kappa values of the experts and trainees were 0.383 and 0.129, respectively, indicating fair and slight agreement with the final diagnosis. Logistic regression analysis identified several factors associated with LN discrimination, including short axis, hilum appearance, texture, and growth. The predictive model was formulated as follows: Logit (p)=−1.566+0.955×(short axis≥10 mm=1)+1.442×(absent hilum through EUS=1)+1.064×(heterogeneous texture=1)+0.676×(conglomerations or fusions=1). The Hosmer-Lemeshow test indicated a good fit (p=0.750). Among the 65 malignant cases and 24 of the 40 benign cases, 53 were misdiagnosed (one lesion was misdiagnosed by ResNet-50). Given the lowest sensitivity and specificity in the trainee group, we calculated the diagnostic efficacy of the combination of ResNet-50, an expert group, and a logistic regression model. This increased the sensitivity to 100.0%. However, the specificity and accuracy decreased from 70.0% and 85.7% to 60.0% and 84.8%, respectively. The detailed diagnostic performances of all approaches are presented in Table 6.
Table 6.
Diagnostic ability of all the modalities
| Method | Sensitivity (%) | Specificity (%) | Accuracy (%) | AUC | PPV (%) | NPV (%) | F1 score |
|---|---|---|---|---|---|---|---|
| ResNet-50+experts+logistic regression analysis | 100.0 | 60.0 | 84.8 | 0.800 | 80.2 | 100.0 | 0.89 |
| ResNet-50 | 95.4 | 70.0 | 85.7 | 0.827 | 83.8 | 90.3 | 0.89 |
| Experts | 73.9 | 65.0 | 70.5 | 0.694 | 77.4 | 60.5 | 0.76 |
| Trainees | 58.5 | 55.0 | 57.1 | 0.567 | 67.9 | 44.9 | 0.63 |
| Logistic regression analysis | 81.5 | 60.0 | 73.3 | 0.708 | 76.8 | 66.7 | 0.79 |
AUC, area under the curve; PPV, positive predictive value; NPV, negative predictive value.
Paired comparison of AUC
The ResNet-50 held the best AUC, whether combined with expert expertise and logistic regression analysis or not (p>0.05). In addition, no significant difference was observed between the logistic regression model and expert expertise (p>0.05). However, both yielded significantly higher AUC values than the trainees (p<0.01). Table 7 shows the differences in AUC among all diagnostic modalities.
Table 7.
Paired comparison of ROC curves
| Pair | Difference of AUC (95% CI) | p-value |
|---|---|---|
| Combination versus ResNet-50 | −0.027 (−0.081 to 0.027) | 0.325 |
| Combination versus logistic regression analysis | 0.092 (0.045−0.140) | 0.000 |
| Combination versus experts | 0.106 (0.042−0.170) | 0.001 |
| Combination versus trainees | 0.233 (0.155−0.310) | 0.000 |
| ResNet-50 versus logistic regression analysis | 0.119 (0.052−0.186) | 0.000 |
| ResNet-50 versus experts | 0.133 (0.065−0.201) | 0.000 |
| ResNet-50 versus trainees | 0.260 (0.178−0.341) | 0.000 |
| Logistic regression analysis versus experts | 0.013 (−0.047 to 0.074) | 0.661 |
| Logistic regression analysis versus trainees | 0.140 (0.069−0.212) | 0.000 |
| Experts versus trainees | 0.127 (0.053−0.200) | 0.001 |
ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval.
DISCUSSION
The accurate identification of lymphadenopathy, particularly in cases of multiple lesions, is crucial. Despite the availability of EUS-FNA/B, the selection of target LNs remains challenging. Although the characteristics differentiating benign and malignant LNs through EUS have been explored, a consensus on the diagnostic criteria has not yet been established. In the present study, we introduced an ML model for diagnosing mediastinal and intra-abdominal lymphadenopathy based on gastrointestinal EUS and demonstrated its potential for aiding discrimination.
Given the imbalance between benign and malignant LNs in the validation dataset, the F1 score was selected as the primary outcome because of its capacity to integrate precision and recall. ResNet-50 achieved the highest F1 score and accuracy, indicating its potential as a complementary tool for human diagnosis. Unexpectedly, it outperformed other ML models with fewer layers. This finding is partly attributable to limited sample size and inconsistent image labels. The latter is unavoidable because of the interference from structures analogous to LNs, such as blood vessels.
To our knowledge, indicators of malignancy may include a short axis ≥10 mm, round shape, sharp margins, absence of hilum, heterogeneous texture, aggregation, and fusion.3,6,16,17 The logistic regression analysis revealed that LNs exhibiting the following characteristics tended to be malignant: short axis ≥10 mm, absent hilum, heterogeneous texture, and aggregation or fusion. Owing to the considerable heterogeneity in LN morphology, we categorized them into two groups based on their shape regularity. The analysis indicated no significant morphological differences between benign and malignant LNs. This conclusion may be related to classification, as we considered LNs with circular, oval, triangular, and reniform shapes as regular lesions. Moreover, although some studies hold that malignant LNs often present with distinct margins because of tumor invasion, our study did not substantiate this opinion. According to Ryu et al., an indistinct margin may be observed in the presence of an extracapsular extension.18 In the present study, the performance of the logistic regression model was not inferior to that of the expert group (p>0.05), suggesting that the model possessed good quality and reference values. ResNet-50 outperformed the logistic regression model established based on expert expertise in terms of sensitivity, specificity, and accuracy, with a significantly higher AUC (p<0.001). Thus, ML could provide clinical decision-making support.
The combination of ResNet-50, logistic regression, and expert analysis achieved optimal sensitivity, underscoring the necessity of collaborative intelligence despite no significant improvement in the F1 score or AUC compared to ResNet-50 alone (p>0.05). A significant difference in the AUC between experts and trainees (p<0.01) suggested the value of clinical expertise. Furthermore, ResNet-50 achieved an optimal independent AUC, indicating its potential for aiding LN diagnosis in settings with limited specialist access.
Although our findings contribute to this field, some limitations must be noted. The generalizability of our results requires further validation given the single-center, retrospective nature of our study. The small sample size constrained the ML development and introduced bias and overfitting. Variations in equipment, EUS frequency, parameter settings, and patient factors could affect the grayscale images, potentially undermining the efficacy of the ML model. In addition, ResNet-50 was trained on static images rather than videos, thus casting uncertainty in its real-time application. Our findings did not apply to tuberculous lesions as patients with tuberculosis were excluded because of poor image quality. EUS based on B-mode imaging is not reliable for the differential diagnosis between benign and malignant LNs, although several image enhancement techniques (such as EUS-elastography and contrast-enhanced harmonic EUS) increase the diagnostic ability.19,20 Nevertheless, considering the increased computational complexity and reduction in the generalization ability of the model, we excluded color images from our training process. In addition, several modifications to the U-Net have been developed, including advances in attention mechanisms, multiscale training, and feature fusion.21 Exploration of these aspects was lacking in this study. The integration of ML should be subject to ongoing monitoring and quality control to ensure patient privacy. With enhanced transparency and interpretability, ML models can gain wider public acceptance, thereby facilitating their application in clinical practice.
In conclusion, ML may be a valuable tool for predicting mediastinal and intra-abdominal LNs based on gastrointestinal EUS images, particularly when combined with expert expertise and logistic regression analysis. However, further studies with larger sample sizes and more rigorous methodologies are necessary to confirm these findings.
Footnotes
Conflicts of Interest
The authors have no potential conflicts of interest.
Funding
This study was supported by the State Key Laboratory of Radiation Medicine and Protection, Soochow University, Suzhou 215123, China (GZK1202402); the Suzhou Multicenter Clinical Research Project for Major Diseases (DZXYJ202404); and the Project of Suzhou Science and Technology Development Plan (Applied Basic Research—Medical and Health Technology Innovation) (SYW2024002).
Author Contributions
Conceptualization: FZJ, GZY, HDM; Data curation: FZJ, GZY, BR, LQK, WW, XLM; Methodology: FZJ, GZY, LJB, LXZ; Formal analysis: FZJ, LJB, LXZ; Project administration: CGL, HDM; Supervision: CGL, HDM; Funding acquisition: HDM; Writing–original draft: FZJ, GZY; Writing–review & editing: all authors.
REFERENCES
- 1.Yasuda I, Tsurumi H, Omar S, et al. Endoscopic ultrasound-guided fine-needle aspiration biopsy for lymphadenopathy of unknown origin. Endoscopy. 2006;38:919–924. doi: 10.1055/s-2006-944665. [DOI] [PubMed] [Google Scholar]
- 2.Malikowski T, Levy MJ, Gleeson FC, et al. Endoscopic ultrasound/fine needle aspiration is effective for lymph node staging in patients with cholangiocarcinoma. Hepatology. 2020;72:940–948. doi: 10.1002/hep.31077. [DOI] [PubMed] [Google Scholar]
- 3.Catalano MF, Sivak MV, Jr, Rice T, et al. Endosonographic features predictive of lymph node metastasis. Gastrointest Endosc. 1994;40:442–446. doi: 10.1016/s0016-5107(94)70206-3. [DOI] [PubMed] [Google Scholar]
- 4.Gill KR, Ghabril MS, Jamil LH, et al. Endosonographic features predictive of malignancy in mediastinal lymph nodes in patients with lung cancer. Gastrointest Endosc. 2010;72:265–271. doi: 10.1016/j.gie.2010.02.037. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Jenssen C, Dietrich CF. Endoscopic ultrasound-guided fine-needle aspiration biopsy and trucut biopsy in gastroenterology: an overview. Best Pract Res Clin Gastroenterol. 2009;23:743–759. doi: 10.1016/j.bpg.2009.05.006. [DOI] [PubMed] [Google Scholar]
- 6.Gleeson FC, Clain JE, Papachristou GI, et al. Prospective assessment of EUS criteria for lymphadenopathy associated with rectal cancer. Gastrointest Endosc. 2009;69:896–903. doi: 10.1016/j.gie.2008.04.051. [DOI] [PubMed] [Google Scholar]
- 7.Chen L, Li Y, Gao X, et al. High diagnostic accuracy and safety of endoscopic ultrasound-guided fine-needle aspiration in malignant lymph nodes: a systematic review and meta-analysis. Dig Dis Sci. 2021;66:2763–2775. doi: 10.1007/s10620-020-06554-2. [DOI] [PubMed] [Google Scholar]
- 8.Li J, Zhi X, Chen J, et al. Deep learning with convex probe endobronchial ultrasound multimodal imaging: a validated tool for automated intrathoracic lymph nodes diagnosis. Endosc Ultrasound. 2021;10:361–371. doi: 10.4103/eus-d-20-00207. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Churchill IF, Gatti AA, Hylton DA, et al. An artificial intelligence algorithm to predict nodal metastasis in lung cancer. Ann Thorac Surg. 2022;114:248–256. doi: 10.1016/j.athoracsur.2021.06.082. [DOI] [PubMed] [Google Scholar]
- 10.Kim YH, Kim GH, Kim KB, et al. Application of a convolutional neural network in the diagnosis of gastric mesenchymal tumors on endoscopic ultrasonography images. J Clin Med. 2020;9:3162. doi: 10.3390/jcm9103162. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Seven G, Silahtaroglu G, Seven OO, et al. Differentiating gastrointestinal stromal tumors from leiomyomas using a neural network trained on endoscopic ultrasonography images. Dig Dis. 2022;40:427–435. doi: 10.1159/000520032. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Minoda Y, Ihara E, Komori K, et al. Efficacy of endoscopic ultrasound with artificial intelligence for the diagnosis of gastrointestinal stromal tumors. J Gastroenterol. 2020;55:1119–1126. doi: 10.1007/s00535-020-01725-4. [DOI] [PubMed] [Google Scholar]
- 13.Shao H, McCarthy C, Wehrenberg-Klee E, et al. CT-guided percutaneous needle biopsy of retroperitoneal and pelvic lymphadenopathy: assessment of technique, diagnostic yield, and clinical value. J Vasc Interv Radiol. 2018;29:1429–1436. doi: 10.1016/j.jvir.2018.03.028. [DOI] [PubMed] [Google Scholar]
- 14.Azad R, Aghdam EK, Rauland A, et al. Medical image segmentation review: the success of U-Net. IEEE Trans Pattern Anal Mach Intell. 2024;46:10076–10095. doi: 10.1109/tpami.2024.3435571. [DOI] [PubMed] [Google Scholar]
- 15.Li K, Xu Y, Zhao Z, et al. Automatic recognition of abdominal organs in ultrasound images based on deep neural networks and K-nearest-neighbor classification. 2021 IEEE International Conference on Robotics and Biomimetics (ROBIO); 2021 Dec 27-31; Sanya, China. pp. 1980–1985. [DOI] [Google Scholar]
- 16.Jamil LH, Kashani A, Scimeca D, et al. Can endoscopic ultrasound distinguish between mediastinal benign lymph nodes and those involved by sarcoidosis, lymphoma, or metastasis? Dig Dis Sci. 2014;59:2191–2198. doi: 10.1007/s10620-014-3164-9. [DOI] [PubMed] [Google Scholar]
- 17.Lin CK, Chang LY, Yu KL, et al. Differentiating metastatic lymph nodes in lung cancer patients based on endobronchial ultrasonography features. Med Ultrason. 2018;20:154–158. doi: 10.11152/mu-1282. [DOI] [PubMed] [Google Scholar]
- 18.Ryu KH, Lee KH, Ryu J, et al. Cervical lymph node imaging reporting and data system for ultrasound of cervical lymphadenopathy: a pilot study. AJR Am J Roentgenol. 2016;206:1286–1291. doi: 10.2214/ajr.15.15381. [DOI] [PubMed] [Google Scholar]
- 19.Tamanini G, Cominardi A, Brighi N, et al. Endoscopic ultrasound assessment and tissue acquisition of mediastinal and abdominal lymph nodes. World J Gastrointest Oncol. 2021;13:1475–1491. doi: 10.4251/wjgo.v13.i10.1475. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Xu W, Shi J, Zeng X, et al. EUS elastography for the differentiation of benign and malignant lymph nodes: a meta-analysis. Gastrointest Endosc. 2011;74:1001–1009. doi: 10.1016/j.gie.2011.07.026. [DOI] [PubMed] [Google Scholar]
- 21.Schlemper J, Oktay O, Schaap M, et al. Attention gated networks: learning to leverage salient regions in medical images. Med Image Anal. 2019;53:197–207. doi: 10.1016/j.media.2019.01.012. [DOI] [PMC free article] [PubMed] [Google Scholar]