Developing a deep learning model for predicting ovarian cancer in Ovarian-Adnexal Reporting and Data System Ultrasound (O-RADS US) Category 4 lesions: A multicenter study

Developing a deep learning model for predicting ovarian cancer in Ovarian-Adnexal Reporting and Data System Ultrasound (O-RADS US) Category 4 lesions: A multicenter study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Developing a deep learning model for predicting ovarian cancer in Ovarian-Adnexal Reporting and Data System Ultrasound (O-RADS US) Category 4 lesions: A multicenter study Wenting Xie, Wenjie Lin, Ping Li, Hongwei Lai, Zhilan Wang, Peizhong Liu, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4457256/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 7 You are reading this latest preprint version Abstract Purpose To develop a deep learning (DL) model for differentiating between benign and malignant ovarian tumors of Ovarian-Adnexal Reporting and Data System Ultrasound (O-RADS US) Category 4 lesions, and validate its diagnostic performance. Methods A retrospective analysis of 1619 US images obtained from three centers from December 2014 to March 2023. DeepLabV3 and YOLOv8 were jointly used to segment, classify, and detect ovarian tumors. Precision and recall and area under the receiver operating characteristic curve (AUC) were employed to assess the model performance. Results A total of 519 patients (including 269 benign and 250 malignant masses) were enrolled in the study. The number of women included in the training, validation, and test cohorts was 426, 46, and 47, respectively. The detection models exhibited an average precision of 98.68% (95% CI: 0.95–0.99) for benign masses and 96.23% (95% CI: 0.92–0.98) for malignant masses. Moreover, in the training set, the AUC was 0.96 (95% CI: 0.94–0.97), whereas in the validation set, the AUC was 0.93(95% CI: 0.89–0.94) and 0.95 (95% CI: 0.91–0.96) in the test set. The sensitivity, specificity, accuracy, positive predictive value, and negative predictive values for the training set were 0.943,0.957,0.951,0.966, and 0.936, respectively, whereas those for the validation set were 0.905,0.935, 0.935,0.919, and 0.931, respectively. In addition, the sensitivity, specificity, accuracy, positive predictive value, and negative predictive value for the test set were 0.925, 0.955, 0.941, 0.956, and 0.927, respectively. Conclusion The constructed DL model exhibited high diagnostic performance in distinguishing benign and malignant ovarian tumors in O-RADS US category 4 lesions. Ovarian cancer Deep learning Ultrasonography Ovarian-Adnexal Reporting and Data System Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Ovarian cancer is one of the most common gynecological malignancy and the fifth leading cause of cancer-related deaths in women worldwide(Siegel et al., 2022 ; Zheng et al., 2020 ). The cancer lacks typical symptoms which makes it difficult to conduct early screening and timely diagnosis. Consequently, most patients with ovarian cancer are diagnosed at an advanced stage(Terp et al., 2023 ). Advanced ovarian cancer patients are often treated with debulking surgery combined with platinum and paclitaxel chemotherapy, but still these treatment modalities are associated with poor survival and high recurrence rates(Konstantinopoulos and Matulonis, 2023 ; Wheeler et al., 2023 ). Currently, the main diagnostic method for ovarian-adnexal lesions is pelvic imaging(Sadowski et al., 2023 ). Ultrasonography (US) is the most commonly used imaging modality for assessing ovarian-adnexal lesions owing to its universality, non-invasive nature, and affordability(Wang et al., 2021 ). However, given its pathological diversity and morphological complexity, accurate preoperative diagnosis of ovarian-adnexal masses by conventional US alone has been suboptimal(Shi et al., 2023 ). To date, several US-based models have been developed to assess the benign or malignant adnexal masses, such as the International Ovarian Tumor Analysis (IOTA) Simple Rules, the Assessment of Different Neoplasia in the Adnexa (ADNEX) model, and Ovarian-Adnexal Reporting and Data System (O-RADS)(Dang Thi Minh et al., 2024 ; Pelayo et al., 2023 ; Pozzati et al., 2023 ). In 2020, the American College of Radiology proposed the O-RADS risk stratification and management system which provides a detailed description of each category(Andreotti et al., 2018 ; Andreotti et al., 2020 ; Yang et al., 2023 ). The system recommends six risk classification categories. These include: O-RADS category 0, defined as an incomplete evaluation; O-RADS category 1, defined as the physiologic category; O-RADS category 2, defined as the almost certainly benign category with <1% malignant probability; O-RADS category 3, referring to lesions with 1% to<10% risk of malignancy; O-RADS category 4, defined as lesions with 10% to<50% risk of malignancy and O-RADS category 5, referring to lesions with high risk of malignancy (≥ 50%)(Vara et al., 2022 ). The accuracy of gynecologic ultrasonography is largely dependent on the sonologist's subjective assessment. It has been observed that the correct classification of the adnexal lesions in expert ultrasound examination is higher than in less experienced doctors(Wu et al., 2023 ). Currently, there are no effective management strategies for O-RADS US 4 lesions, with the risk of malignancy exhibiting significant variations and some lesions found to be benign. If benign lesions can be accurately diagnosed, patients can avoid unnecessary or extensive surgery. A simple description of the ovarian tumor by a sonologist may not be completely fulfilled. This calls for the establishment of appropriate and advanced approaches for sub-stratifying O-RADS US 4 lesions into benign and malignant subgroups. Artificial intelligence (AI) has emerged as a significant tool with diverse medical applications(Wang et al., 2024 ). Deep learning (DL) can quantitatively analyze of medical images and has been applied in the field of oncology (Taddese et al., 2024 ). Several studies demonstrated that DL can improve the diagnosis, predict treatment responses, and progression-free survival of patients with ovarian tumors(Arezzo et al., 2022 ; Boehm et al., 2022 ; Na et al., 2024 ; Sadeghi et al., 2024 ; Yao et al., 2021 ). Compared with traditional imaging diagnosis by radiologists, the DL method can improve the accuracy and reduce the bias of diagnosis results(Chen et al., 2022 ). However, few studies have demonstrated the diagnostic performance of AI in O-RADS 4 lesions. In this multicenter study, we developed and tested a US image-based DL model for distinguishing between benign and malignant ovarian tumors of O-RADS 4 lesions and to demonstrate the diagnostic performance of our DL model. Materials and Methods Study population and datasets This retrospective study was approved by the Institutional Review Board of the Second Affiliated Hospital of Fujian Medical University (No.636, 2023). The analyzed datasets were obtained from three hospitals from December 2014 to March 2023. The three centers were coded as A, the Second Affiliated Hospital of Fujian Medical University; B, Fujian Cancer Hospital; and C, Nanping First Hospital Affiliated to Fujian Medical University. Consecutive patients who met the following criteria were included: (1) underwent diagnostic pelvis US prior to gynecological surgery; (2) diagnosed with O-RADS category 4 by US radiologists according to the O-RADS lexicon white paper; (3) postoperative pathological confirmation of ovarian tumor; (4) did not use radiotherapy or chemotherapy before US examination. Patients with poor image quality and without histopathology were excluded. A total of 519 women met the inclusion criteria and were enrolled in this study (Fig. 1 ). Center C was used as the validation set, 10% of all cases coded Center A and Center B as the test set, and the remaining 90% as the training set. The US image target masses with the largest diameter or a more complicated ultrasound morphology were selected for further analysis. The US images were acquired using ultrasound devices equipped with abdominal probes spanning frequencies from 1 to 6 MHz, as well as transvaginal probes covering frequencies from 2 to 9 MHz. Confirmation of all adnexal lesions was achieved by surgical pathology. Furthermore, pertinent clinical data such as age at diagnosis, lesion size, CA125 levels, and menopausal status were meticulously documented for subsequent analysis. Algorithm for analysis Image annotation First, we used a professional image annotation tool LabelImg ( https://github.com/HumanSignal/labelImg ) to draw bounding boxes on all ovarian cancer images to mark the location and extent of ovarian tumors. The boundaries of each lesion area were drawn by two radiologists (with 6 and 3 years of experience). The boundaries of all lesions were manually delineated on the axes of the 2D images. In addition, we accurately outlined the tumor borders, ensuring precise adjustment of the bounding box to accommodate the morphology of the tumor. In addition to the bounding box labeling, each bounding box was assigned a corresponding category label, benign or malignant. Feature extraction The dataset was preprocessed which involved processed such as image specification and data enhancement. Images were then resized to the same resolution. In addition, we randomly employed data enhancement techniques such as flipping, cropping, and addition of noise to the images to generate richer training data to improve the generalization of the model. To normalize the range of values for each phenotypic feature, the feature normalization technique was employed to convert these features to 0 or 1. Next, segmentation feature extraction was conducted using DeepLabV3(Chen et al., 2017 ). DeepLabV3 utilizes the architecture of domain-adaptive convolutional neural networks to capture the regional features of objects. Within the encoder-decoder architecture, the encoder module employs ResNet-101(Jusman, 2023) as its backbone network for extracting low-level features. Subsequently, the decoder module utilizes null convolution to augment edge features, followed by up-sampling to reconstruct the segmentation output. In addition, we used the YOLOv8 (Terven et al., 2023 ) network to extract the target detection features. YOLOv8 adopts Darknet-53(Azim et al., 2022) as the backbone network to improve the inter-channel correlation of the feature channels while preserving the fine edge information due to the new channel attention module and short connections. The output feature maps were post-processed to predict the edge coordinates and the classification. Machine learning model development Furthermore, we propose a novel dual-model architecture (Fig. 2 ) that leverages the complementary abilities of DeepLabV3 and YOLOv8. In our approach, we employed DeepLabV3 to first perform pixel-accurate segmentation of ovarian tumor regions from US images. The ResNet-101pretrained on ImageNet was employed as the encoder backbone in DeepLabV3. Atrous separable convolutions were utilized with varying rates to explicitly capture multi-scale information. Moreover, a lightweight decoder module was integrated to restore fine spatial details and generate segmentation maps at the full image resolution. These segmented outputs were subsequently fed into YOLOv8 for downstream classification and detection tasks. Notably, YOLOv8 features a modified Darknet-53 network, which consists of residual and convolutional blocks, serving as the backbone for feature extraction. YOLOv8 classifies each tumor region as benign or malignant and generates bounding boxes around individual tumor objects. This enables joint analysis of segmentation, classification, and localization within a single framework. To achieve better outcomes, we first trained DeepLabV3 on our US image datasets until convergence was attained. The model outputs were then extracted and fed to YOLOv8 in the training phase. Both models were fine-tuned end-to-end through this sequential process. At inference, a new US image is directly input to obtain segmented tumor regions, which are instantly classified and localized by YOLOv8. Model training A pipeline was established to extract tumor ROIs at 512x512 pixels from the annotated masks while maintaining original aspect ratios. To address the class imbalance, we oversampled the minority malignant class during the training stage. We trained DeepLabV3 and YOLOv8 sequentially in an end-to-end fashion on our dataset through transfer learning. DeepLabV3 was first pretrained on PASCAL VOC3 to achieve semantic segmentation. Subsequently, we fine-tuned the model on our ovarian tumor masks for 50 epochs. This process employed stochastic gradient descent with polynomial learning rate decay, utilizing a batch size of 4 and employing the dice coefficient loss function. After segmentation, YOLOv8 was initialized from weights pretrained on COCO4 for object detection. It was then optimized on the extracted DeepLabV3 tumor regions for 100 epochs. The proposed dual-model architecture and training procedure presents an innovative approach for joint analysis of US images. Statistical analysis All statistical analyses were conducted using the SPSS 20.0 software (IBM, Armonk, NY, USA). Categorical data were analyzed by Chi-Squared test and expressed as the frequency and percentage. Continuous data were analyzed by the Student’s t test or Mann-Whitney U test and then expressed as the mean and standard deviation or median and interquartile range. The ROC curve analysis was utilized to evaluate the performance of our proposed dual-model method. The ROC curves were constructed by the ROC package on Python (version 3.8.0). p < 0.05 was considered statistically significant. Results Patient characteristics A total of 1619 images in 519 patients who underwent US examination and surgery were enrolled from three hospitals. The histological profile of enrolled adnexal patients is presented in Table 1 . The number of women included in the training, validation, and test cohorts was 426, 46, and 47, respectively. Table 1 Pathology results of 519 adnexal masses that were assigned O-RADS US category 4. Center A (n = 156) Center B (n = 317) Center C (n = 46) Histopathological findings No. of patients (%) Histopathological findings No. of patients (%) Histopathological findings No. of patients (%) Benign Benign Benign Benign cyst 5(3.2) Benign cyst 13(4.1) Serous cyst 2(4.3) Endometriosis cyst 6(3.8) Endometriosis cyst 21(6.6) Endometriosis cyst 6(13.0) Teratoma 8(5.1) Teratoma 17(5.4) Teratoma 5(10.9) Cystadenoma 24(15.4) Cystadenofibroma 7(2.2) Cystadenoma 22(47.8) Struma ovarii 1(0.6) Cystadenoma 61(19.2) Struma ovarii 1(2.2) Inflammation 5(3.2) Struma ovarii 10(3.2) Fibroma 1(2.2) Cystadenofibroma 4(2.6) Brenner 2(0.6) Sclerosing stromal tumor 1(2.2) Fibroma 6(3.8) Inflammation 5(1.6) Theca-fibroma 1(0.6) Fibroma 4(1.3) Theca-fibroma 30(9.5) Microcystic stromal tumour 1(0.3) Malignant Malignant Malignant Borderline 46(29.5) Borderline 71(22.4) Borderline 6(13.0) High grade 33(21.2) High grade 32(10.1) High grade 1(2.2) Immature teratoma 5(3.2) Low grade 2(0.6) Metastasis 1(2.2) Clear cell carcinoma 5(3.2) Immature teratoma 3(1.0) Endometrioid cancer 3(2.0) Clear cell carcinoma 14(4.4) Granular cell tumor 3(2.0) Endometrioid cancer 5(1.6) Sertoli-Leydig cell tumor 1(0.6) Granular cell tumor 9(2.8) Sertoli-Leydig cell tumor 1(0.3) Metastasis 8(2.5) Malignant mixed Mullerian tumour 1(0.3) Center A,the second affiliated hospital of fujian medical university; Center B, Fujian Cancer Hospital; Center C, Nanping First Hospital Affiliated to Fujian Medical University. The clinical characteristics and laboratory results of patients are shown in Table 2 . It can be inferred that 269 (51.8%) lesions were benign while 250 (48.2%) lesions were malignant. Among the study variables, the largest diameter of the lesions (mm) (108 (75.5-142.6) vs. 120.5 (77.7-171.7) mm, p = 0.024) was significantly different between benign and malignant groups. In women with elevated CA125 levels, malignant tumors were significantly more prevalent than benign tumors ( p < 0.001). Nevertheless, no statistically significant variances were detected in terms of age at diagnosis, menopausal status, or tumor location between the benign and malignant groups. Table 2 Comparison clinical features between benign and malignant O-RADS 4 US adnexal lesions. Characteristic Benign (n = 269) Malignant (n = 250) p-value Age at diagnosis 47.0 ± 15.9 47.2 ± 14.5 0.866 Largest diameter of lesion (mm) 108(75.5−142.6) 120.5(77.7−171.7) 0.024 Menopausal status 0.631 Premenopausal 145(53.9) 140(56.0) Postmenopausal 124(46.1) 110(44.0) Location 0.38 Left 126(46.8) 118(47.2) Right 119(44.2) 101(40.4) Bilateral 24(9.0) 31(12.4) Serum CA125 level (U/ml) 0.000 ≤ 35 176(65.4) 105(42.0) >35 93(34.6) 145(58.0) Model performance The performance of the detection DL model is shown in Fig. 3 . The average precision of the precision-recall curve was 98.68% (95% CI: 0.95–0.99) for benign masses (Fig. 3 A) and 96.23% (95% CI: 0.92–0.98) for malignant masses (Fig. 3 B). When applied to US imaging of adnexal masses, the model showed the potential to identify nodules in benign and malignant categories (Fig. 4 ). The DL model had the best discrimination between the benign and malignant groups, with an AUC of 0.96 (95% CI: 0.94–0.97) in the training set, an AUC of 0.93(95% CI: 0.89–0.94) in the validation set and 0.95 (95% CI: 0.91–0.96) in the test set (Fig. 5 ). Further analysis indicated that the sensitivity, specificity, accuracy, positive predictive value, and negative predictive value in the training set were 0.943, 0.957, 0.951, 0.966, and 0.936, respectively, whereas those for the validation set were 0.905,0.935, 0.935,0.919, and 0.931, respectively. In addition, the sensitivity, specificity, accuracy, positive predictive value, and negative predictive value for the test set were 0.925,0.955,0.941,0.956, and 0.927, respectively (Table 3 ). Table 3 Diagnostic performance of the deep learning classification model in training, validation, and test cohorts. Set AUC Sensitivity Specificity Accuracy Positive predictive value Negative predictive value Training set 0.96(95%CI:0.94–0.97) 0.943 0.957 0.951 0.966 0.936 Validation set 0.93(95%CI:0.89–0.94) 0.905 0.935 0.927 0.919 0.931 Test set 0.95(95%CI:0.91–0.96) 0.925 0.955 0.941 0.956 0.927 Discussion The significant similarity in ultrasonographic characteristics between malignant and benign ovarian lesions posed a diagnostic challenge for sonologists. Nevertheless, prior research has demonstrated that O-RADS US can accurately detect ovarian malignancies, indicating outstanding diagnostic accuracy(Hack et al., 2022 ). However, the risk of malignancy in O-RADS 4 was in the range of 10% to<50% implying that some benign lesions were classified in this category. Correct classification of an adnexal lesion is therefore important for improving personalized management. Xu et al . incorporated the qualitative parameters of contrast-enhanced ultrasound (CEUS) to reassign the O-RADS category and the overall sensitivity increased to 90.2%(Xu et al., 2023 ). In this study, we evaluated the performance of the DL model to classify benign or malignant lesions in O-RADS US Category 4 lesions and showed an acceptable diagnostic performance with an AUC of 0.95. We deployed a DL method for distinguishing O-RADS 4 lesion. DeepLabV3 is well-suited for precise semantic segmentation due to its powerful encoder-decoder design. Chen et al . developed DeepLabV3 to facilitate the segmentation of colon cancer histology at subcellular scales(Chen et al., 2018 ). However, semantic segmentation alone does not provide classification or localization of tumors. Object detection models such as YOLO have achieved great success in natural images. In a previous study, Xiao et al . employed YOLOv3 to identify lung cancer in CT scans (Xiao et al., 2023 ). Meanwhile, YOLOv8 has been shown to rapidly classify and achieve object detection due to its robust architecture(Redmon J). In this study, we leveraged the strengths of these models to conduct an in-depth analysis of US images. Our innovative dual-model architecture merges the features of DeepLabV3 and YOLOv8, enabling concurrent segmentation, classification, and detection of ovarian tumors. DeepLabV3 first segments tumor regions, whose outputs are then classified and localized by YOLOv8. This novel dual-model architecture can enable efficient and accurate analysis of imaging analysis. Artificial intelligence has been demonstrated to improve the diagnostics rate of ovarian tumors. A study found that using deep neural networks to analyze ultrasound images can discriminate between benign and malignant ovarian masses and achieve comparable diagnostic accuracy to expert examiners(Christiansen et al., 2021 ). A recent multicenter study developed a deep convolutional neural network (DCNN) model for detecting ovarian cancer with high performance(Gao et al., 2022 ). By integrating DeepLabV3 and YOLOv8, our DL system achieved a higher AUC. However, we cannot directly compare our results to those obtained in previous studies because our focus was on a specific population subset, and the ovarian tumor datasets utilized differ from those in previous investigations. In addition, we investigated the clinical features of benign and malignant O-RADS 4 US lesions. The levels of CA-125 levels were higher in malignant masses than in benign masses ( p <0.05), which is consistent with recent literature(Wong et al., 2023 ). Not surprisingly, the largest diameter of the lesion was significantly higher in malignant masses than in benign masses ( p = 0.024). This finding may be ascribed to the concept that ovarian tumors often lack typical symptoms and hence detected at the advanced stage. This study has several limitations that should be acknowledged. First, this study was a retrospective investigation which may have inherent biases. To improve the DL model’s reliability, a prospective study should be performed. Secondly, we did not include clinical factors in the DL model. Third, we only used one DL method to construct the model. In our future work, we plan to explore and compare the performance of various DL methods. Moreover, we did not assess the diagnostic accuracy of our DL model against US-based models, such as the IOTA-ADNEX model. Therefore, it will be imperative to conduct a comparative analysis of diagnostic performance between our model and the IOTA-ADNEX model. Conclusions In conclusion, this study demonstrates that the US image-based DL model may be used as a tool for distinguishing between benign and malignant ovarian tumors of O-RADS 4 lesions. Abbreviations O-RADS, Ovarian-Adnexal Reporting and Data System US,Ultrasonography DL,Deep learning AUC, Area under the receiver operating characteristic curve Declarations Author Contribution WTX and WJL were major contributors in writing the manuscript. WTX, WJL, PL,HWL,ZLW and YJH analysis the images and collected and organized the data. PZL and YL analysis the deep learning model. LNT an GRLdesigned this study. WTX and WJL contributed equally. All authors read and approved the fnal manuscript. Declarations of interest None References Andreotti, R.F., Timmerman, D., Benacerraf, B.R., Bennett, G.L., Bourne, T., Brown, D.L., Coleman, B.G., Frates, M.C., Froyman, W., Goldstein, S.R., et al. (2018). Ovarian-Adnexal Reporting Lexicon for Ultrasound: A White Paper of the ACR Ovarian-Adnexal Reporting and Data System Committee. Journal of the American College of Radiology : JACR 15, 1415-1429. https://doi.org/10.1016/j.jacr.2018.07.004. Andreotti, R.F., Timmerman, D., Strachowski, L.M., Froyman, W., Benacerraf, B.R., Bennett, G.L., Bourne, T., Brown, D.L., Coleman, B.G., Frates, M.C., et al. (2020). O-RADS US Risk Stratification and Management System: A Consensus Guideline from the ACR Ovarian-Adnexal Reporting and Data System Committee. Radiology 294, 168-185. https://doi.org/10.1148/radiol.2019191150. Arezzo, F., Cormio, G., La Forgia, D., Santarsiero, C.M., Mongelli, M., Lombardi, C., Cazzato, G., Cicinelli, E., and Loizzi, V. (2022). A machine learning approach applied to gynecological ultrasound to predict progression-free survival in ovarian cancer patients. Arch Gynecol Obstet 306, 2143-2154. https://doi.org/10.1007/s00404-022-06578-1. Azim,T.(2022). Breast Cancer Identification Using Improved DarkNet53 Model. In Innovations in Bio-Inspired Computing and Applications: Proceedings of the 13th International Conference on Innovations in Bio-Inspired Computing and Applications (IBICA 2022) Held During December 15-17 (Vol. 649, p. 338). Boehm, K.M., Aherne, E.A., Ellenson, L., Nikolovski, I., Alghamdi, M., Vázquez-García, I., Zamarin, D., Long Roche, K., Liu, Y., Patel, D., et al. (2022). Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer. Nature Cancer 3, 723-733.https://doi.org/10.1038/s43018-022-00388-9. Chen, H., Yang, B.W., Qian, L., Meng, Y.S., Bai, X.H., Hong, X.W., He, X., Jiang, M.J., Yuan, F., Du, Q.W., et al. (2022).Deep Learning Prediction of Ovarian Malignancy at US Compared with O-RADS and Expert Assessment. Radiology 304,106-113. https://doi.org/10.1148/radiol.211367. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., and Yuille, A.L. (2018). DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. IEEE transactions on pattern analysis and machine intelligence 40, 834-848. https://doi.org/10.1109/TPAMI.2017.2699184. Chen,L. C., Papandreou,G., Schroff, F., Adam,H. (2017). Rethinking atrous convolution for semantic image segmentation. arXiv preprint, arXiv:1706.05587. Christiansen, F., Epstein, E.L., Smedberg, E., Åkerlund, M., Smith, K., and Epstein, E. (2021). Ultrasound image analysis using deep neural networks for discriminating between benign and malignant ovarian tumors: comparison with expert subjective assessment. Ultrasound in Obstetrics & Gynecology 57, 155-163. https://doi.org/10.1002/uog.23530. Dang Thi Minh, N., Nguyen Van, T., Duong Duc, H., Nguyen Tuan, M., Duong Thi Tra, G., Do Tuan, D., and Nguyen Tai, D. (2024). IOTA simple rules: An efficient tool for evaluation of ovarian tumors by non-experienced but trained examiners - A prospective study. Heliyon 10, e24262. https://doi.org/ 10.1016/j.heliyon.2024.e24262. Gao, Y., Zeng, S., Xu, X., Li, H., Yao, S., Song, K., Li, X., Chen, L., Tang, J., Xing, H., et al. (2022). Deep learning-enabled pelvic ultrasound images for accurate diagnosis of ovarian cancer in China: a retrospective, multicentre, diagnostic study. The Lancet Digital Health 4, e179-e187. https://doi.org/10.1016/S2589-7500(21)00278-8. Hack, K., Gandhi, N., Bouchard-Fortier, G., Chawla, T.P., Ferguson, S.E., Li, S., Kahn, D., Tyrrell, P.N., and Glanc, P. (2022). External Validation of O-RADS US Risk Stratification and Management System. Radiology 304, 114-120. https://doi.org/10.1148/radiol.211868. Jusman,Y. Comparison of Prostate Cell Image Classification Using CNN: ResNet-101 and VGG-19. In 2023 IEEE 13th International Conference on Control System, Computing and Engineering (ICCSCE),2023. pp.74-78. Konstantinopoulos, P.A., and Matulonis, U.A. (2023). Clinical and translational advances in ovarian cancer therapy. Nat Cancer 4, 1239-1257. https://doi.org/10.1038/s43018-023-00617-9. Na, I., Noh, J.J., Kim, C.K., Lee, J.W., and Park, H. (2024). Combined radiomics-clinical model to predict platinum-sensitivity in advanced high-grade serous ovarian carcinoma using multimodal MRI. Front Oncol 14, 1341228. https://doi.org/10.3389/fonc.2024.1341228. Pelayo, M., Sancho-Sauco, J., Sánchez-Zurdo, J., Perez-Mies, B., Abarca-Martínez, L., Cancelo-Hidalgo, M.J., Sainz-Bueno, J.A., Alcázar, J.L., and Pelayo-Delgado, I. (2023). Application of Ultrasound Scores (Subjective Assessment, Simple Rules Risk Assessment, ADNEX Model, O-RADS) to Adnexal Masses of Difficult Classification. Diagnostics (Basel, Switzerland) 13:2785. https://doi.org/10.3390/diagnostics13172785. Pozzati, F., Sassu, C.M., Marini, G., Mascilini, F., Biscione, A., Giannarelli, D., Garganese, G., Fragomeni, S.M., Scambia, G., Testa, A.C., et al. (2023). Subjective assessment and IOTA ADNEX model in evaluation of adnexal masses in patients with history of breast cancer. Ultrasound Obstet Gynecol 62, 594-602. https://doi.org/ 10.1002/uog.26253. Redmon J, F.A. YOLOv3: An Incremental Improvement. arXiv. 2018;arXiv:1804.02767. Sadeghi, M.H., Sina, S., Omidi, H., Farshchitabrizi, A.H., and Alavi, M. (2024). Deep learning in ovarian cancer diagnosis: a comprehensive review of various imaging modalities. Polish Journal of Radiology 89, 30-48. https://doi.org/ 10.5114/pjr.2024.134817. Sadowski, E.A., Rockall, A., Thomassin-Naggara, I., Barroilhet, L.M., Wallace, S.K., Jha, P., Gupta, A., Shinagare, A.B., Guo, Y., and Reinhold, C. (2023). Adnexal Lesion Imaging: Past, Present, and Future. Radiology 307:e223281. https://doi.org/ 10.1148/radiol.223281. Shi, Y., Li, H., Wu, X., Li, X., and Yang, M. (2023). O-RADS combined with contrast-enhanced ultrasound in risk stratification of adnexal masses. Journal of ovarian research 16:153. https://doi.org/10.1186/s13048-023-01243-w. Siegel, R.L., Miller, K.D., Fuchs, H.E., and Jemal, A. (2022). Cancer statistics, 2022. CA: A Cancer Journal for Clinicians 72, 7-33. https://doi.org/10.3322/caac.21708. Taddese, A.A., Tilahun, B.C., Awoke, T., Atnafu, A., Mamuye, A., and Mengiste, S.A. (2024). Deep-learning models for image-based gynecological cancer diagnosis: a systematic review and meta-analysis. Frontiers in Oncology 13:1216326. https://doi.org/10.3389/fonc.2023.1216326. Terp, S.K., Stoico, M.P., Dybkær, K., and Pedersen, I.S. (2023). Early diagnosis of ovarian cancer based on methylation profiles in peripheral blood cell-free DNA: a systematic review. Clinical Epigenetics 15:24. https://doi.org/ 10.1186/s13148-023-01440-w. Terven,J., Córdova-Esparza,D.M., Romero-González,J. A.(2023). A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas. Machine Learning and Knowledge Extraction, 5(4)1680-1716. Vara, J., Manzour, N., Chacón, E., López-Picazo, A., Linares, M., Pascual, M.Á., Guerriero, S., and Alcázar, J.L. (2022). Ovarian Adnexal Reporting Data System (O-RADS) for Classifying Adnexal Masses: A Systematic Review and Meta-Analysis. Cancers 14, 3151. https://doi.org/ 10.3390/cancers14133151. Wang, H., Liu, C., Zhao, Z., Zhang, C., Wang, X., Li, H., Wu, H., Liu, X., Li, C., Qi, L., et al. (2021). Application of Deep Convolutional Neural Networks for Discriminating Benign, Borderline, and Malignant Serous Ovarian Tumors From Ultrasound Images. Front Oncol 11, 770683. https://doi.org/ 10.3389/fonc.2021.770683. Wang, Y., Lin, W., Zhuang, X., Wang, X., He, Y., Li, L., and Lyu, G. (2024). Advances in artificial intelligence for the diagnosis and treatment of ovarian cancer (Review). Oncology reports 51:46. https://doi.org/ 10.3892/or.2024.8705. Wheeler, V., Umstead, B., and Chadwick, C. (2023). Adnexal Masses: Diagnosis and Management. American family physician 108, 580-587. Wong, B.Z.Y., Causa Andrieu, P.I., Sonoda, Y., Chi, D.S., Aviki, E.M., Vargas, H.A., and Woo, S. (2023). Improving risk stratification of indeterminate adnexal masses on MRI: What imaging features help predict malignancy in O-RADS MRI 4 lesions? European journal of radiology 168, 111122. https://doi.org/10.1016/j.ejrad.2023.111122. Wu, M., Zhang, M., Cao, J., Wu, S., Chen, Y., Luo, L., Lin, X., Su, M., and Zhang, X. (2023). Predictive accuracy and reproducibility of the O-RADS US scoring system among sonologists with different training levels. Arch Gynecol Obstet 308, 631-637. https://doi.org/10.1007/s00404-022-06752-5. Xiao, H., Xue, X., Zhu, M., Jiang, X., Xia, Q., Chen, K., Li, H., Long, L., and Peng, K. (2023). Deep learning-based lung image registration: A review. Computers in biology and medicine 165, 107434. https://doi.org/10.1016/j.compbiomed.2023.107434. Xu, J., Huang, Z., Zeng, J., Zheng, Z., Cao, J., Su, M., and Zhang, X. (2023). Value of Contrast-Enhanced Ultrasound Parameters in the Evaluation of Adnexal Masses with Ovarian–Adnexal Reporting and Data System Ultrasound. Ultrasound in medicine & biology 49, 1527-1534. https://doi.org/10.1016/j.ultrasmedbio.2023.02.015. Yang, Y., Wang, H., Liu, Z., Su, N., Gao, L., Tao, X., Zhang, R., Gu, Y., Ma, L., Wang, R., et al. (2023). Effect of differences in O-RADS lexicon interpretation between senior and junior sonologists on O-RADS classification and diagnostic performance. Journal of Cancer Research and Clinical Oncology 149, 12275-12283. https://doi.org/10.1007/s00432-023-05108-z. Yao, F., Ding, J., Hu, Z., Cai, M., Liu, J., Huang, X., Zheng, R., Lin, F., and Lan, L. (2021). Ultrasound-based radiomics score: a potential biomarker for the prediction of progression-free survival in ovarian epithelial cancer. Abdominal Radiology 46, 4936-4945. https://doi.org/ 10.1007/s00261-021-03163-z. Zheng, L., Cui, C., Shi, O., Lu, X., Li, Y.-k., Wang, W., Li, Y., and Wang, Q. (2020). Incidence and mortality of ovarian cancer at the global, regional, and national levels, 1990–2017. Gynecologic oncology 159, 239-247. https://doi.org/10.1016/j.ygyno.2020.07.008. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 06 Jun, 2024 Reviews received at journal 05 Jun, 2024 Reviewers agreed at journal 25 May, 2024 Reviewers invited by journal 23 May, 2024 Editor assigned by journal 22 May, 2024 Submission checks completed at journal 22 May, 2024 First submitted to journal 21 May, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4457256","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":309703975,"identity":"53fd2c8c-d040-4d4a-af76-9d24a2780fe6","order_by":0,"name":"Wenting Xie","email":"","orcid":"","institution":"Department of Ultrasound Medicine, The Second Affiliated Hospital of Fujian medical University","correspondingAuthor":false,"prefix":"","firstName":"Wenting","middleName":"","lastName":"Xie","suffix":""},{"id":309703976,"identity":"327d785d-112d-44a4-bad5-181926a01f7d","order_by":1,"name":"Wenjie Lin","email":"","orcid":"","institution":"Department of Ultrasound Medicine, The Second Affiliated Hospital of Fujian medical University","correspondingAuthor":false,"prefix":"","firstName":"Wenjie","middleName":"","lastName":"Lin","suffix":""},{"id":309703977,"identity":"b5a89fde-6db5-4b67-94ab-f3dc822320cb","order_by":2,"name":"Ping Li","email":"","orcid":"","institution":"Department of Gynecology and Obstetrics, Quanzhou First Hospital Affiliated to Fujian Medical University","correspondingAuthor":false,"prefix":"","firstName":"Ping","middleName":"","lastName":"Li","suffix":""},{"id":309703978,"identity":"da02d46e-8041-4580-98d7-ff881029cff9","order_by":3,"name":"Hongwei Lai","email":"","orcid":"","institution":"Department of Ultrasound, Fujian Provincial Maternity and Children’s Hospital","correspondingAuthor":false,"prefix":"","firstName":"Hongwei","middleName":"","lastName":"Lai","suffix":""},{"id":309703979,"identity":"d10cb0e6-d83f-4b2f-8c3c-58faec35059d","order_by":4,"name":"Zhilan Wang","email":"","orcid":"","institution":"Department of Ultrasound, Nanping First Hospital Affiliated to Fujian Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zhilan","middleName":"","lastName":"Wang","suffix":""},{"id":309703980,"identity":"e73049c5-206b-4e86-892b-1524944665d5","order_by":5,"name":"Peizhong Liu","email":"","orcid":"","institution":"School of Medicine, Huaqiao University","correspondingAuthor":false,"prefix":"","firstName":"Peizhong","middleName":"","lastName":"Liu","suffix":""},{"id":309703981,"identity":"7edbe462-414a-4e65-8487-f56cbeab0568","order_by":6,"name":"Yijun Huang","email":"","orcid":"","institution":"Department of Ultrasound, Clinical Oncology School of Fujian Medical University, Fujian Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yijun","middleName":"","lastName":"Huang","suffix":""},{"id":309703982,"identity":"14508ea0-4200-4ff4-90a6-1e5033bb2c4f","order_by":7,"name":"Yao Liu","email":"","orcid":"","institution":"Quanzhou Bolang Technology Group Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Yao","middleName":"","lastName":"Liu","suffix":""},{"id":309703983,"identity":"326bd857-143e-42cf-99cf-70a12c28bb88","order_by":8,"name":"Lina Tang","email":"","orcid":"","institution":"Department of Ultrasound, Clinical Oncology School of Fujian Medical University, Fujian Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Lina","middleName":"","lastName":"Tang","suffix":""},{"id":309703984,"identity":"ec84063f-1c64-492f-b5f2-1ed9909390d2","order_by":9,"name":"Guorong Lyu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAwklEQVRIiWNgGAWjYBACPhBRAcT8DIwNxGlhAxFngFiygWQtBgeIdRgb/9oHDAcqDtsbnz/c9uAHg52cLiHL2CSeGzAcOHM4cduNxHbDHoZkYzNC1rFJHGNg/th2OMHsBmObBA/DgcRtxGhhOPgP6LD+g22Sf4jSwt8G1NJwmHEDQ2KbNJG2AMPswLH0xBk3gFpkDIjwCz8/0GEHaqzt+fuPP5N8U2EnR1ALg0QC+w8Ez4CQcrA1BA0dBaNgFIyCEQ8ATqk/d27DDo4AAAAASUVORK5CYII=","orcid":"","institution":"Department of Ultrasound Medicine, The Second Affiliated Hospital of Fujian medical University","correspondingAuthor":true,"prefix":"","firstName":"Guorong","middleName":"","lastName":"Lyu","suffix":""}],"badges":[],"createdAt":"2024-05-21 22:53:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4457256/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4457256/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":57942181,"identity":"b748291b-e36a-4e17-97df-ad466d00b3b0","added_by":"auto","created_at":"2024-06-07 19:00:18","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":224537,"visible":true,"origin":"","legend":"\u003cp\u003eThe flowchart of patient enrollment, inclusion, and exclusion criteria, and partitioning of datasets. Center A, the second affiliated hospital of Fujian Medical University; Center B, Fujian Cancer Hospital; Center C, Nanping First Hospital Affiliated to Fujian Medical University.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-4457256/v1/04cb289621919b950d763012.png"},{"id":57942183,"identity":"94cea04c-754a-4016-a54d-3e16b63348a1","added_by":"auto","created_at":"2024-06-07 19:00:19","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":257697,"visible":true,"origin":"","legend":"\u003cp\u003eThe overview of the proposed framework. Our approach is divided into two major phases: 1) The segmentation stage in which the DeepLabV3 model is used to obtain the lesion segmentation mask, which is subsequently used to mask the input image. 2) In the classifier phase, YOLOv8 is employed to disentangle the feature maps of a CNN into human-interpretable concepts.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-4457256/v1/ecbd5a2368f716193eae9bdc.png"},{"id":57942188,"identity":"2f02be0d-9857-4fb3-a695-8bd72500c17e","added_by":"auto","created_at":"2024-06-07 19:00:21","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":12708,"visible":true,"origin":"","legend":"\u003cp\u003eThe performance of the adnexal mass detection model. (A) The precision-recall curve of the detection model used for evaluating for performance of benign masses indicating an average precision (AP) of 0.98 (95% CI: 0.95-0.99). (B) The precision-recall curve of the detection model used for evaluating for performance of malignant masses displaying an AP of 0.96 (95% CI: 0.92-0.98).\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-4457256/v1/bd9c55860f65ffb25bd1771a.png"},{"id":57942184,"identity":"5fd76c19-d9f6-451a-ad60-89b3d92178c3","added_by":"auto","created_at":"2024-06-07 19:00:20","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":757622,"visible":true,"origin":"","legend":"\u003cp\u003eImages with O-RADS US category 4 masses detected using the model. (A) High-grade serous carcinoma in a 70-year-old female patient, (B) Borderline serous tumor in a 57-year-old female patient, (C) Brenner (borderline) in a 69-year-old female patient, (D) Endometriosis cyst in a 22-year-old female patient, (E) Teratoma in a 22-year-old female patient, (F) Mucinous cystadenoma in a 47-year-old female patient. The red and blue bounding box within each imaging indicates the adnex masses delineated by the detection model.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-4457256/v1/4db2dc3ca742449f4865985a.png"},{"id":57942185,"identity":"a7895980-40e8-4d7e-aebb-971606e0001c","added_by":"auto","created_at":"2024-06-07 19:00:20","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":15049,"visible":true,"origin":"","legend":"\u003cp\u003eThe ROC curve showing the classification model for ovarian cancer classification.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-4457256/v1/6f631a2ca3777c9831c46929.png"},{"id":57942223,"identity":"07e79acc-61f7-43d9-b42f-0e5771c7bbec","added_by":"auto","created_at":"2024-06-07 19:00:31","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2120697,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4457256/v1/6a988e43-c6e8-4725-8bf8-cd0eb3ddf602.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Developing a deep learning model for predicting ovarian cancer in Ovarian-Adnexal Reporting and Data System Ultrasound (O-RADS US) Category 4 lesions: A multicenter study","fulltext":[{"header":"Introduction","content":"\u003cp\u003eOvarian cancer is one of the most common gynecological malignancy and the fifth leading cause of cancer-related deaths in women worldwide(Siegel et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Zheng et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). The cancer lacks typical symptoms which makes it difficult to conduct early screening and timely diagnosis. Consequently, most patients with ovarian cancer are diagnosed at an advanced stage(Terp et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Advanced ovarian cancer patients are often treated with debulking surgery combined with platinum and paclitaxel chemotherapy, but still these treatment modalities are associated with poor survival and high recurrence rates(Konstantinopoulos and Matulonis, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Wheeler et al., \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Currently, the main diagnostic method for ovarian-adnexal lesions is pelvic imaging(Sadowski et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Ultrasonography (US) is the most commonly used imaging modality for assessing ovarian-adnexal lesions owing to its universality, non-invasive nature, and affordability(Wang et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). However, given its pathological diversity and morphological complexity, accurate preoperative diagnosis of ovarian-adnexal masses by conventional US alone has been suboptimal(Shi et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). To date, several US-based models have been developed to assess the benign or malignant adnexal masses, such as the International Ovarian Tumor Analysis (IOTA) Simple Rules, the Assessment of Different Neoplasia in the Adnexa (ADNEX) model, and Ovarian-Adnexal Reporting and Data System (O-RADS)(Dang Thi Minh et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Pelayo et al., \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Pozzati et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn 2020, the American College of Radiology proposed the O-RADS risk stratification and management system which provides a detailed description of each category(Andreotti et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Andreotti et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Yang et al., \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). The system recommends six risk classification categories. These include: O-RADS category 0, defined as an incomplete evaluation; O-RADS category 1, defined as the physiologic category; O-RADS category 2, defined as the almost certainly benign category with \u0026lt;1% malignant probability; O-RADS category 3, referring to lesions with 1% to\u0026lt;10% risk of malignancy; O-RADS category 4, defined as lesions with 10% to\u0026lt;50% risk of malignancy and O-RADS category 5, referring to lesions with high risk of malignancy (\u0026ge;\u0026thinsp;50%)(Vara et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). The accuracy of gynecologic ultrasonography is largely dependent on the sonologist's subjective assessment. It has been observed that the correct classification of the adnexal lesions in expert ultrasound examination is higher than in less experienced doctors(Wu et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Currently, there are no effective management strategies for O-RADS US 4 lesions, with the risk of malignancy exhibiting significant variations and some lesions found to be benign. If benign lesions can be accurately diagnosed, patients can avoid unnecessary or extensive surgery. A simple description of the ovarian tumor by a sonologist may not be completely fulfilled. This calls for the establishment of appropriate and advanced approaches for sub-stratifying O-RADS US 4 lesions into benign and malignant subgroups.\u003c/p\u003e \u003cp\u003eArtificial intelligence (AI) has emerged as a significant tool with diverse medical applications(Wang et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Deep learning (DL) can quantitatively analyze of medical images and has been applied in the field of oncology (Taddese et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Several studies demonstrated that DL can improve the diagnosis, predict treatment responses, and progression-free survival of patients with ovarian tumors(Arezzo et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Boehm et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Na et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Sadeghi et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Yao et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Compared with traditional imaging diagnosis by radiologists, the DL method can improve the accuracy and reduce the bias of diagnosis results(Chen et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). However, few studies have demonstrated the diagnostic performance of AI in O-RADS 4 lesions.\u003c/p\u003e \u003cp\u003eIn this multicenter study, we developed and tested a US image-based DL model for distinguishing between benign and malignant ovarian tumors of O-RADS 4 lesions and to demonstrate the diagnostic performance of our DL model.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy population and datasets\u003c/h2\u003e \u003cp\u003e This retrospective study was approved by the Institutional Review Board of the Second Affiliated Hospital of Fujian Medical University (No.636, 2023). The analyzed datasets were obtained from three hospitals from December 2014 to March 2023. The three centers were coded as A, the Second Affiliated Hospital of Fujian Medical University; B, Fujian Cancer Hospital; and C, Nanping First Hospital Affiliated to Fujian Medical University. Consecutive patients who met the following criteria were included: (1) underwent diagnostic pelvis US prior to gynecological surgery; (2) diagnosed with O-RADS category 4 by US radiologists according to the O-RADS lexicon white paper; (3) postoperative pathological confirmation of ovarian tumor; (4) did not use radiotherapy or chemotherapy before US examination. Patients with poor image quality and without histopathology were excluded. A total of 519 women met the inclusion criteria and were enrolled in this study (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Center C was used as the validation set, 10% of all cases coded Center A and Center B as the test set, and the remaining 90% as the training set.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe US image target masses with the largest diameter or a more complicated ultrasound morphology were selected for further analysis. The US images were acquired using ultrasound devices equipped with abdominal probes spanning frequencies from 1 to 6 MHz, as well as transvaginal probes covering frequencies from 2 to 9 MHz. Confirmation of all adnexal lesions was achieved by surgical pathology. Furthermore, pertinent clinical data such as age at diagnosis, lesion size, CA125 levels, and menopausal status were meticulously documented for subsequent analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eAlgorithm for analysis\u003c/h2\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003eImage annotation\u003c/h2\u003e \u003cp\u003eFirst, we used a professional image annotation tool LabelImg (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/HumanSignal/labelImg\u003c/span\u003e\u003cspan address=\"https://github.com/HumanSignal/labelImg\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) to draw bounding boxes on all ovarian cancer images to mark the location and extent of ovarian tumors. The boundaries of each lesion area were drawn by two radiologists (with 6 and 3 years of experience). The boundaries of all lesions were manually delineated on the axes of the 2D images. In addition, we accurately outlined the tumor borders, ensuring precise adjustment of the bounding box to accommodate the morphology of the tumor. In addition to the bounding box labeling, each bounding box was assigned a corresponding category label, benign or malignant.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eFeature extraction\u003c/h2\u003e \u003cp\u003eThe dataset was preprocessed which involved processed such as image specification and data enhancement. Images were then resized to the same resolution. In addition, we randomly employed data enhancement techniques such as flipping, cropping, and addition of noise to the images to generate richer training data to improve the generalization of the model. To normalize the range of values for each phenotypic feature, the feature normalization technique was employed to convert these features to 0 or 1. Next, segmentation feature extraction was conducted using DeepLabV3(Chen et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). DeepLabV3 utilizes the architecture of domain-adaptive convolutional neural networks to capture the regional features of objects. Within the encoder-decoder architecture, the encoder module employs ResNet-101(Jusman, 2023) as its backbone network for extracting low-level features. Subsequently, the decoder module utilizes null convolution to augment edge features, followed by up-sampling to reconstruct the segmentation output. In addition, we used the YOLOv8 (Terven et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) network to extract the target detection features. YOLOv8 adopts Darknet-53(Azim et al., 2022) as the backbone network to improve the inter-channel correlation of the feature channels while preserving the fine edge information due to the new channel attention module and short connections. The output feature maps were post-processed to predict the edge coordinates and the classification.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eMachine learning model development\u003c/h2\u003e \u003cp\u003eFurthermore, we propose a novel dual-model architecture (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) that leverages the complementary abilities of DeepLabV3 and YOLOv8. In our approach, we employed DeepLabV3 to first perform pixel-accurate segmentation of ovarian tumor regions from US images. The ResNet-101pretrained on ImageNet was employed as the encoder backbone in DeepLabV3. Atrous separable convolutions were utilized with varying rates to explicitly capture multi-scale information. Moreover, a lightweight decoder module was integrated to restore fine spatial details and generate segmentation maps at the full image resolution. These segmented outputs were subsequently fed into YOLOv8 for downstream classification and detection tasks. Notably, YOLOv8 features a modified Darknet-53 network, which consists of residual and convolutional blocks, serving as the backbone for feature extraction. YOLOv8 classifies each tumor region as benign or malignant and generates bounding boxes around individual tumor objects. This enables joint analysis of segmentation, classification, and localization within a single framework. To achieve better outcomes, we first trained DeepLabV3 on our US image datasets until convergence was attained. The model outputs were then extracted and fed to YOLOv8 in the training phase. Both models were fine-tuned end-to-end through this sequential process. At inference, a new US image is directly input to obtain segmented tumor regions, which are instantly classified and localized by YOLOv8.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eModel training\u003c/h2\u003e \u003cp\u003eA pipeline was established to extract tumor ROIs at 512x512 pixels from the annotated masks while maintaining original aspect ratios. To address the class imbalance, we oversampled the minority malignant class during the training stage. We trained DeepLabV3 and YOLOv8 sequentially in an end-to-end fashion on our dataset through transfer learning. DeepLabV3 was first pretrained on PASCAL VOC3 to achieve semantic segmentation. Subsequently, we fine-tuned the model on our ovarian tumor masks for 50 epochs. This process employed stochastic gradient descent with polynomial learning rate decay, utilizing a batch size of 4 and employing the dice coefficient loss function. After segmentation, YOLOv8 was initialized from weights pretrained on COCO4 for object detection. It was then optimized on the extracted DeepLabV3 tumor regions for 100 epochs. The proposed dual-model architecture and training procedure presents an innovative approach for joint analysis of US images.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eAll statistical analyses were conducted using the SPSS 20.0 software (IBM, Armonk, NY, USA). Categorical data were analyzed by Chi-Squared test and expressed as the frequency and percentage. Continuous data were analyzed by the Student\u0026rsquo;s \u003cem\u003et\u003c/em\u003e test or Mann-Whitney \u003cem\u003eU\u003c/em\u003e test and then expressed as the mean and standard deviation or median and interquartile range. The ROC curve analysis was utilized to evaluate the performance of our proposed dual-model method. The ROC curves were constructed by the ROC package on Python (version 3.8.0). p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003ePatient characteristics\u003c/h2\u003e \u003cp\u003eA total of 1619 images in 519 patients who underwent US examination and surgery were enrolled from three hospitals. The histological profile of enrolled adnexal patients is presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The number of women included in the training, validation, and test cohorts was 426, 46, and 47, respectively.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePathology results of 519 adnexal masses that were assigned O-RADS US category 4.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eCenter A\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;156)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e \u003cp\u003eCenter B\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;317)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003eCenter C\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;46)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHistopathological findings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNo. of patients (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHistopathological findings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNo. of patients (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHistopathological findings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo. of patients (%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eBenign\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eBenign\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003eBenign\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBenign cyst\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5(3.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBenign cyst\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13(4.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSerous cyst\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2(4.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEndometriosis cyst\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6(3.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEndometriosis cyst\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e21(6.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eEndometriosis cyst\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e6(13.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTeratoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8(5.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTeratoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e17(5.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTeratoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5(10.9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCystadenoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e24(15.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCystadenofibroma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7(2.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCystadenoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e22(47.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStruma ovarii\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1(0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCystadenoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e61(19.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eStruma ovarii\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1(2.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInflammation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5(3.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStruma ovarii\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10(3.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eFibroma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1(2.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCystadenofibroma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4(2.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBrenner\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2(0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSclerosing stromal tumor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1(2.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFibroma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6(3.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eInflammation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5(1.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTheca-fibroma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1(0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFibroma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4(1.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTheca-fibroma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e30(9.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMicrocystic stromal tumour\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1(0.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMalignant\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eMalignant\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003eMalignant\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBorderline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e46(29.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBorderline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e71(22.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBorderline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e6(13.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHigh grade\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e33(21.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh grade\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e32(10.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHigh grade\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1(2.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eImmature teratoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5(3.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow grade\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2(0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMetastasis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1(2.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClear cell carcinoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5(3.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eImmature teratoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3(1.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEndometrioid cancer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3(2.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eClear cell carcinoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14(4.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGranular cell tumor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3(2.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEndometrioid cancer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5(1.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSertoli-Leydig cell tumor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1(0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGranular cell tumor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9(2.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSertoli-Leydig cell tumor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1(0.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMetastasis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8(2.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMalignant mixed Mullerian tumour\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1(0.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003eCenter A,the second affiliated hospital of fujian medical university; Center B, Fujian Cancer Hospital; Center C, Nanping First Hospital Affiliated to Fujian Medical University.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe clinical characteristics and laboratory results of patients are shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. It can be inferred that 269 (51.8%) lesions were benign while 250 (48.2%) lesions were malignant. Among the study variables, the largest diameter of the lesions (mm) (108 (75.5-142.6) vs. 120.5 (77.7-171.7) mm, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.024) was significantly different between benign and malignant groups. In women with elevated CA125 levels, malignant tumors were significantly more prevalent than benign tumors (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Nevertheless, no statistically significant variances were detected in terms of age at diagnosis, menopausal status, or tumor location between the benign and malignant groups.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison clinical features between benign and malignant O-RADS 4 US adnexal lesions.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCharacteristic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBenign\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;269)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMalignant\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;250)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge at diagnosis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e47.0\u0026thinsp;\u0026plusmn;\u0026thinsp;15.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e47.2\u0026thinsp;\u0026plusmn;\u0026thinsp;14.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.866\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLargest diameter of lesion (mm)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e108(75.5\u0026minus;142.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e120.5(77.7\u0026minus;171.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.024\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMenopausal status\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.631\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePremenopausal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e145(53.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e140(56.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePostmenopausal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e124(46.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e110(44.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLocation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.38\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLeft\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e126(46.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e118(47.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e119(44.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e101(40.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBilateral\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e24(9.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e31(12.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSerum CA125 level (U/ml)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026le;\u0026thinsp;35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e176(65.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e105(42.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026gt;35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e93(34.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e145(58.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eModel performance\u003c/h2\u003e \u003cp\u003eThe performance of the detection DL model is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The average precision of the precision-recall curve was 98.68% (95% CI: 0.95\u0026ndash;0.99) for benign masses (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA) and 96.23% (95% CI: 0.92\u0026ndash;0.98) for malignant masses (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). When applied to US imaging of adnexal masses, the model showed the potential to identify nodules in benign and malignant categories (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe DL model had the best discrimination between the benign and malignant groups, with an AUC of 0.96 (95% CI: 0.94\u0026ndash;0.97) in the training set, an AUC of 0.93(95% CI: 0.89\u0026ndash;0.94) in the validation set and 0.95 (95% CI: 0.91\u0026ndash;0.96) in the test set (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). Further analysis indicated that the sensitivity, specificity, accuracy, positive predictive value, and negative predictive value in the training set were 0.943, 0.957, 0.951, 0.966, and 0.936, respectively, whereas those for the validation set were 0.905,0.935, 0.935,0.919, and 0.931, respectively. In addition, the sensitivity, specificity, accuracy, positive predictive value, and negative predictive value for the test set were 0.925,0.955,0.941,0.956, and 0.927, respectively (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDiagnostic performance of the deep learning classification model in training, validation, and test cohorts.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSet\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePositive predictive value\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNegative predictive value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTraining set\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.96(95%CI:0.94\u0026ndash;0.97)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.943\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.957\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.951\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.966\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.936\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValidation set\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.93(95%CI:0.89\u0026ndash;0.94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.905\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.935\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.927\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.919\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.931\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTest set\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.95(95%CI:0.91\u0026ndash;0.96)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.925\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.955\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.941\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.956\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.927\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe significant similarity in ultrasonographic characteristics between malignant and benign ovarian lesions posed a diagnostic challenge for sonologists. Nevertheless, prior research has demonstrated that O-RADS US can accurately detect ovarian malignancies, indicating outstanding diagnostic accuracy(Hack et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). However, the risk of malignancy in O-RADS 4 was in the range of 10% to\u0026lt;50% implying that some benign lesions were classified in this category. Correct classification of an adnexal lesion is therefore important for improving personalized management. Xu \u003cem\u003eet al\u003c/em\u003e. incorporated the qualitative parameters of contrast-enhanced ultrasound (CEUS) to reassign the O-RADS category and the overall sensitivity increased to 90.2%(Xu et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). In this study, we evaluated the performance of the DL model to classify benign or malignant lesions in O-RADS US Category 4 lesions and showed an acceptable diagnostic performance with an AUC of 0.95.\u003c/p\u003e \u003cp\u003eWe deployed a DL method for distinguishing O-RADS 4 lesion. DeepLabV3 is well-suited for precise semantic segmentation due to its powerful encoder-decoder design. Chen \u003cem\u003eet al\u003c/em\u003e. developed DeepLabV3 to facilitate the segmentation of colon cancer histology at subcellular scales(Chen et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). However, semantic segmentation alone does not provide classification or localization of tumors. Object detection models such as YOLO have achieved great success in natural images. In a previous study, Xiao \u003cem\u003eet al\u003c/em\u003e. employed YOLOv3 to identify lung cancer in CT scans (Xiao et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Meanwhile, YOLOv8 has been shown to rapidly classify and achieve object detection due to its robust architecture(Redmon J). In this study, we leveraged the strengths of these models to conduct an in-depth analysis of US images. Our innovative dual-model architecture merges the features of DeepLabV3 and YOLOv8, enabling concurrent segmentation, classification, and detection of ovarian tumors. DeepLabV3 first segments tumor regions, whose outputs are then classified and localized by YOLOv8. This novel dual-model architecture can enable efficient and accurate analysis of imaging analysis.\u003c/p\u003e \u003cp\u003eArtificial intelligence has been demonstrated to improve the diagnostics rate of ovarian tumors. A study found that using deep neural networks to analyze ultrasound images can discriminate between benign and malignant ovarian masses and achieve comparable diagnostic accuracy to expert examiners(Christiansen et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). A recent multicenter study developed a deep convolutional neural network (DCNN) model for detecting ovarian cancer with high performance(Gao et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). By integrating DeepLabV3 and YOLOv8, our DL system achieved a higher AUC. However, we cannot directly compare our results to those obtained in previous studies because our focus was on a specific population subset, and the ovarian tumor datasets utilized differ from those in previous investigations.\u003c/p\u003e \u003cp\u003eIn addition, we investigated the clinical features of benign and malignant O-RADS 4 US lesions. The levels of CA-125 levels were higher in malignant masses than in benign masses (\u003cem\u003ep\u003c/em\u003e\u0026lt;0.05), which is consistent with recent literature(Wong et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Not surprisingly, the largest diameter of the lesion was significantly higher in malignant masses than in benign masses (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.024). This finding may be ascribed to the concept that ovarian tumors often lack typical symptoms and hence detected at the advanced stage.\u003c/p\u003e \u003cp\u003eThis study has several limitations that should be acknowledged. First, this study was a retrospective investigation which may have inherent biases. To improve the DL model\u0026rsquo;s reliability, a prospective study should be performed. Secondly, we did not include clinical factors in the DL model. Third, we only used one DL method to construct the model. In our future work, we plan to explore and compare the performance of various DL methods. Moreover, we did not assess the diagnostic accuracy of our DL model against US-based models, such as the IOTA-ADNEX model. Therefore, it will be imperative to conduct a comparative analysis of diagnostic performance between our model and the IOTA-ADNEX model.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eIn conclusion, this study demonstrates that the US image-based DL model may be used as a tool for distinguishing between benign and malignant ovarian tumors of O-RADS 4 lesions.\u003c/p\u003e "},{"header":"Abbreviations","content":"\u003cp\u003eO-RADS, Ovarian-Adnexal Reporting and Data System\u003c/p\u003e\n\u003cp\u003eUS,Ultrasonography\u003c/p\u003e\n\u003cp\u003eDL,Deep learning\u003c/p\u003e\n\u003cp\u003eAUC, Area under the receiver operating characteristic curve\u0026nbsp;\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eWTX and WJL were major contributors in writing the manuscript. WTX, WJL, PL,HWL,ZLW and YJH analysis the images and collected and organized the data. PZL and YL analysis the deep learning model. LNT an GRLdesigned this study. WTX and WJL contributed equally. All authors read and approved the fnal manuscript.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eDeclarations of interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNone\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAndreotti, R.F., Timmerman, D., Benacerraf, B.R., Bennett, G.L., Bourne, T., Brown, D.L., Coleman, B.G., Frates, M.C., Froyman, W., Goldstein, S.R., et al. (2018). Ovarian-Adnexal Reporting Lexicon for Ultrasound: A White Paper of the ACR Ovarian-Adnexal Reporting and Data System Committee. Journal of the American College of Radiology : JACR 15, 1415-1429. https://doi.org/10.1016/j.jacr.2018.07.004. \u003c/li\u003e\n\u003cli\u003eAndreotti, R.F., Timmerman, D., Strachowski, L.M., Froyman, W., Benacerraf, B.R., Bennett, G.L., Bourne, T., Brown, D.L., Coleman, B.G., Frates, M.C., et al. (2020). O-RADS US Risk Stratification and Management System: A Consensus Guideline from the ACR Ovarian-Adnexal Reporting and Data System Committee. Radiology 294, 168-185. https://doi.org/10.1148/radiol.2019191150. \u003c/li\u003e\n\u003cli\u003eArezzo, F., Cormio, G., La Forgia, D., Santarsiero, C.M., Mongelli, M., Lombardi, C., Cazzato, G., Cicinelli, E., and Loizzi, V. (2022). A machine learning approach applied to gynecological ultrasound to predict progression-free survival in ovarian cancer patients. Arch Gynecol Obstet 306, 2143-2154. https://doi.org/10.1007/s00404-022-06578-1. \u003c/li\u003e\n\u003cli\u003eAzim,T.(2022). Breast Cancer Identification Using Improved DarkNet53 Model. In Innovations in Bio-Inspired Computing and Applications: Proceedings of the 13th International Conference on Innovations in Bio-Inspired Computing and Applications (IBICA 2022) Held During December 15-17 (Vol. 649, p. 338). \u003c/li\u003e\n\u003cli\u003eBoehm, K.M., Aherne, E.A., Ellenson, L., Nikolovski, I., Alghamdi, M., V\u0026aacute;zquez-Garc\u0026iacute;a, I., Zamarin, D., Long Roche, K., Liu, Y., Patel, D., et al. (2022). Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer. Nature Cancer 3, 723-733.https://doi.org/10.1038/s43018-022-00388-9. \u003c/li\u003e\n\u003cli\u003eChen, H., Yang, B.W., Qian, L., Meng, Y.S., Bai, X.H., Hong, X.W., He, X., Jiang, M.J., Yuan, F., Du, Q.W., et al. (2022).Deep Learning Prediction of Ovarian Malignancy at US Compared with O-RADS and Expert Assessment. Radiology 304,106-113. https://doi.org/10.1148/radiol.211367. \u003c/li\u003e\n\u003cli\u003eChen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., and Yuille, A.L. (2018). DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. IEEE transactions on pattern analysis and machine intelligence 40, 834-848. https://doi.org/10.1109/TPAMI.2017.2699184. \u003c/li\u003e\n\u003cli\u003eChen,L. C., Papandreou,G., Schroff, F., Adam,H. (2017). Rethinking atrous convolution for semantic image segmentation. arXiv preprint, arXiv:1706.05587.\u003c/li\u003e\n\u003cli\u003eChristiansen, F., Epstein, E.L., Smedberg, E., \u0026Aring;kerlund, M., Smith, K., and Epstein, E. (2021). Ultrasound image analysis using deep neural networks for discriminating between benign and malignant ovarian tumors: comparison with expert subjective assessment. Ultrasound in Obstetrics \u0026amp; Gynecology 57, 155-163. https://doi.org/10.1002/uog.23530. \u003c/li\u003e\n\u003cli\u003eDang Thi Minh, N., Nguyen Van, T., Duong Duc, H., Nguyen Tuan, M., Duong Thi Tra, G., Do Tuan, D., and Nguyen Tai, D. (2024). IOTA simple rules: An efficient tool for evaluation of ovarian tumors by non-experienced but trained examiners - A prospective study. Heliyon 10, e24262. https://doi.org/ 10.1016/j.heliyon.2024.e24262. \u003c/li\u003e\n\u003cli\u003eGao, Y., Zeng, S., Xu, X., Li, H., Yao, S., Song, K., Li, X., Chen, L., Tang, J., Xing, H., et al. (2022). Deep learning-enabled pelvic ultrasound images for accurate diagnosis of ovarian cancer in China: a retrospective, multicentre, diagnostic study. The Lancet Digital Health 4, e179-e187. https://doi.org/10.1016/S2589-7500(21)00278-8. \u003c/li\u003e\n\u003cli\u003eHack, K., Gandhi, N., Bouchard-Fortier, G., Chawla, T.P., Ferguson, S.E., Li, S., Kahn, D., Tyrrell, P.N., and Glanc, P. (2022). External Validation of O-RADS US Risk Stratification and Management System. Radiology 304, 114-120. https://doi.org/10.1148/radiol.211868. \u003c/li\u003e\n\u003cli\u003eJusman,Y. Comparison of Prostate Cell Image Classification Using CNN: ResNet-101 and VGG-19. In 2023 IEEE 13th International Conference on Control System, Computing and Engineering (ICCSCE),2023. pp.74-78. \u003c/li\u003e\n\u003cli\u003eKonstantinopoulos, P.A., and Matulonis, U.A. (2023). Clinical and translational advances in ovarian cancer therapy. Nat Cancer 4, 1239-1257. https://doi.org/10.1038/s43018-023-00617-9. \u003c/li\u003e\n\u003cli\u003eNa, I., Noh, J.J., Kim, C.K., Lee, J.W., and Park, H. (2024). Combined radiomics-clinical model to predict platinum-sensitivity in advanced high-grade serous ovarian carcinoma using multimodal MRI. Front Oncol 14, 1341228. https://doi.org/10.3389/fonc.2024.1341228. \u003c/li\u003e\n\u003cli\u003ePelayo, M., Sancho-Sauco, J., S\u0026aacute;nchez-Zurdo, J., Perez-Mies, B., Abarca-Mart\u0026iacute;nez, L., Cancelo-Hidalgo, M.J., Sainz-Bueno, J.A., Alc\u0026aacute;zar, J.L., and Pelayo-Delgado, I. (2023). Application of Ultrasound Scores (Subjective Assessment, Simple Rules Risk Assessment, ADNEX Model, O-RADS) to Adnexal Masses of Difficult Classification. Diagnostics (Basel, Switzerland) 13:2785. https://doi.org/10.3390/diagnostics13172785. \u003c/li\u003e\n\u003cli\u003ePozzati, F., Sassu, C.M., Marini, G., Mascilini, F., Biscione, A., Giannarelli, D., Garganese, G., Fragomeni, S.M., Scambia, G., Testa, A.C., et al. (2023). Subjective assessment and IOTA ADNEX model in evaluation of adnexal masses in patients with history of breast cancer. Ultrasound Obstet Gynecol 62, 594-602. https://doi.org/ 10.1002/uog.26253. \u003c/li\u003e\n\u003cli\u003eRedmon J, F.A. YOLOv3: An Incremental Improvement. arXiv. 2018;arXiv:1804.02767. \u003c/li\u003e\n\u003cli\u003eSadeghi, M.H., Sina, S., Omidi, H., Farshchitabrizi, A.H., and Alavi, M. (2024). Deep learning in ovarian cancer diagnosis: a comprehensive review of various imaging modalities. Polish Journal of Radiology 89, 30-48. https://doi.org/ 10.5114/pjr.2024.134817. \u003c/li\u003e\n\u003cli\u003eSadowski, E.A., Rockall, A., Thomassin-Naggara, I., Barroilhet, L.M., Wallace, S.K., Jha, P., Gupta, A., Shinagare, A.B., Guo, Y., and Reinhold, C. (2023). Adnexal Lesion Imaging: Past, Present, and Future. Radiology 307:e223281. https://doi.org/ 10.1148/radiol.223281. \u003c/li\u003e\n\u003cli\u003eShi, Y., Li, H., Wu, X., Li, X., and Yang, M. (2023). O-RADS combined with contrast-enhanced ultrasound in risk stratification of adnexal masses. Journal of ovarian research 16:153. https://doi.org/10.1186/s13048-023-01243-w. \u003c/li\u003e\n\u003cli\u003eSiegel, R.L., Miller, K.D., Fuchs, H.E., and Jemal, A. (2022). Cancer statistics, 2022. CA: A Cancer Journal for Clinicians 72, 7-33. https://doi.org/10.3322/caac.21708. \u003c/li\u003e\n\u003cli\u003eTaddese, A.A., Tilahun, B.C., Awoke, T., Atnafu, A., Mamuye, A., and Mengiste, S.A. (2024). Deep-learning models for image-based gynecological cancer diagnosis: a systematic review and meta-analysis. Frontiers in Oncology 13:1216326. https://doi.org/10.3389/fonc.2023.1216326. \u003c/li\u003e\n\u003cli\u003eTerp, S.K., Stoico, M.P., Dybk\u0026aelig;r, K., and Pedersen, I.S. (2023). Early diagnosis of ovarian cancer based on methylation profiles in peripheral blood cell-free DNA: a systematic review. Clinical Epigenetics 15:24. https://doi.org/ 10.1186/s13148-023-01440-w. \u003c/li\u003e\n\u003cli\u003eTerven,J., C\u0026oacute;rdova-Esparza,D.M., Romero-Gonz\u0026aacute;lez,J. A.(2023). A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas. Machine Learning and Knowledge Extraction, 5(4)1680-1716.\u003c/li\u003e\n\u003cli\u003eVara, J., Manzour, N., Chac\u0026oacute;n, E., L\u0026oacute;pez-Picazo, A., Linares, M., Pascual, M.\u0026Aacute;., Guerriero, S., and Alc\u0026aacute;zar, J.L. (2022). Ovarian Adnexal Reporting Data System (O-RADS) for Classifying Adnexal Masses: A Systematic Review and Meta-Analysis. Cancers 14, 3151. https://doi.org/ 10.3390/cancers14133151. \u003c/li\u003e\n\u003cli\u003eWang, H., Liu, C., Zhao, Z., Zhang, C., Wang, X., Li, H., Wu, H., Liu, X., Li, C., Qi, L., et al. (2021). Application of Deep Convolutional Neural Networks for Discriminating Benign, Borderline, and Malignant Serous Ovarian Tumors From Ultrasound Images. Front Oncol 11, 770683. https://doi.org/ 10.3389/fonc.2021.770683. \u003c/li\u003e\n\u003cli\u003eWang, Y., Lin, W., Zhuang, X., Wang, X., He, Y., Li, L., and Lyu, G. (2024). Advances in artificial intelligence for the diagnosis and treatment of ovarian cancer (Review). Oncology reports 51:46. https://doi.org/ 10.3892/or.2024.8705. \u003c/li\u003e\n\u003cli\u003eWheeler, V., Umstead, B., and Chadwick, C. (2023). Adnexal Masses: Diagnosis and Management. American family physician 108, 580-587.\u003c/li\u003e\n\u003cli\u003eWong, B.Z.Y., Causa Andrieu, P.I., Sonoda, Y., Chi, D.S., Aviki, E.M., Vargas, H.A., and Woo, S. (2023). Improving risk stratification of indeterminate adnexal masses on MRI: What imaging features help predict malignancy in O-RADS MRI 4 lesions? European journal of radiology 168, 111122. https://doi.org/10.1016/j.ejrad.2023.111122. \u003c/li\u003e\n\u003cli\u003eWu, M., Zhang, M., Cao, J., Wu, S., Chen, Y., Luo, L., Lin, X., Su, M., and Zhang, X. (2023). Predictive accuracy and reproducibility of the O-RADS US scoring system among sonologists with different training levels. Arch Gynecol Obstet 308, 631-637. https://doi.org/10.1007/s00404-022-06752-5. \u003c/li\u003e\n\u003cli\u003eXiao, H., Xue, X., Zhu, M., Jiang, X., Xia, Q., Chen, K., Li, H., Long, L., and Peng, K. (2023). Deep learning-based lung image registration: A review. Computers in biology and medicine 165, 107434. https://doi.org/10.1016/j.compbiomed.2023.107434. \u003c/li\u003e\n\u003cli\u003eXu, J., Huang, Z., Zeng, J., Zheng, Z., Cao, J., Su, M., and Zhang, X. (2023). Value of Contrast-Enhanced Ultrasound Parameters in the Evaluation of Adnexal Masses with Ovarian\u0026ndash;Adnexal Reporting and Data System Ultrasound. Ultrasound in medicine \u0026amp; biology 49, 1527-1534. https://doi.org/10.1016/j.ultrasmedbio.2023.02.015. \u003c/li\u003e\n\u003cli\u003eYang, Y., Wang, H., Liu, Z., Su, N., Gao, L., Tao, X., Zhang, R., Gu, Y., Ma, L., Wang, R., et al. (2023). Effect of differences in O-RADS lexicon interpretation between senior and junior sonologists on O-RADS classification and diagnostic performance. Journal of Cancer Research and Clinical Oncology 149, 12275-12283. https://doi.org/10.1007/s00432-023-05108-z. \u003c/li\u003e\n\u003cli\u003eYao, F., Ding, J., Hu, Z., Cai, M., Liu, J., Huang, X., Zheng, R., Lin, F., and Lan, L. (2021). Ultrasound-based radiomics score: a potential biomarker for the prediction of progression-free survival in ovarian epithelial cancer. Abdominal Radiology 46, 4936-4945. https://doi.org/ 10.1007/s00261-021-03163-z. \u003c/li\u003e\n\u003cli\u003eZheng, L., Cui, C., Shi, O., Lu, X., Li, Y.-k., Wang, W., Li, Y., and Wang, Q. (2020). Incidence and mortality of ovarian cancer at the global, regional, and national levels, 1990\u0026ndash;2017. Gynecologic oncology 159, 239-247. https://doi.org/10.1016/j.ygyno.2020.07.008. \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-cancer-research-and-clinical-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jocr","sideBox":"Learn more about [Journal of Cancer Research and Clinical Oncology](https://www.springer.com/journal/432)","snPcode":"432","submissionUrl":"https://submission.nature.com/new-submission/432/3","title":"Journal of Cancer Research and Clinical Oncology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Ovarian cancer, Deep learning, Ultrasonography, Ovarian-Adnexal Reporting and Data System","lastPublishedDoi":"10.21203/rs.3.rs-4457256/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4457256/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e \u003cb\u003ePurpose\u003c/b\u003e To develop a deep learning (DL) model for differentiating between benign and malignant ovarian tumors of Ovarian-Adnexal Reporting and Data System Ultrasound (O-RADS US) Category 4 lesions, and validate its diagnostic performance.\u003c/p\u003e \u003cp\u003e \u003cb\u003eMethods\u003c/b\u003e A retrospective analysis of 1619 US images obtained from three centers from December 2014 to March 2023. DeepLabV3 and YOLOv8 were jointly used to segment, classify, and detect ovarian tumors. Precision and recall and area under the receiver operating characteristic curve (AUC) were employed to assess the model performance.\u003c/p\u003e \u003cp\u003eResults\u003c/p\u003e \u003cp\u003eA total of 519 patients (including 269 benign and 250 malignant masses) were enrolled in the study. The number of women included in the training, validation, and test cohorts was 426, 46, and 47, respectively. The detection models exhibited an average precision of 98.68% (95% CI: 0.95\u0026ndash;0.99) for benign masses and 96.23% (95% CI: 0.92\u0026ndash;0.98) for malignant masses. Moreover, in the training set, the AUC was 0.96 (95% CI: 0.94\u0026ndash;0.97), whereas in the validation set, the AUC was 0.93(95% CI: 0.89\u0026ndash;0.94) and 0.95 (95% CI: 0.91\u0026ndash;0.96) in the test set. The sensitivity, specificity, accuracy, positive predictive value, and negative predictive values for the training set were 0.943,0.957,0.951,0.966, and 0.936, respectively, whereas those for the validation set were 0.905,0.935, 0.935,0.919, and 0.931, respectively. In addition, the sensitivity, specificity, accuracy, positive predictive value, and negative predictive value for the test set were 0.925, 0.955, 0.941, 0.956, and 0.927, respectively.\u003c/p\u003e \u003cp\u003eConclusion\u003c/p\u003e \u003cp\u003eThe constructed DL model exhibited high diagnostic performance in distinguishing benign and malignant ovarian tumors in O-RADS US category 4 lesions.\u003c/p\u003e","manuscriptTitle":"Developing a deep learning model for predicting ovarian cancer in Ovarian-Adnexal Reporting and Data System Ultrasound (O-RADS US) Category 4 lesions: A multicenter study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-06-07 19:00:00","doi":"10.21203/rs.3.rs-4457256/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-06-06T19:36:47+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-06-05T11:22:47+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"139015216963445375400039133229167500882","date":"2024-05-25T23:22:55+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-05-23T08:13:16+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-05-22T16:01:56+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-05-22T14:36:17+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Cancer Research and Clinical Oncology","date":"2024-05-21T22:42:22+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-cancer-research-and-clinical-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jocr","sideBox":"Learn more about [Journal of Cancer Research and Clinical Oncology](https://www.springer.com/journal/432)","snPcode":"432","submissionUrl":"https://submission.nature.com/new-submission/432/3","title":"Journal of Cancer Research and Clinical Oncology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"d5363187-e091-4442-b1a7-db6f45695745","owner":[],"postedDate":"June 7th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2024-06-27T08:21:13+00:00","versionOfRecord":[],"versionCreatedAt":"2024-06-07 19:00:00","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4457256","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4457256","identity":"rs-4457256","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}