Integrating ultrasound-CT-MR for preoperative multi-task prediction in ovarian cancer: achieving diagnostic parity with multidisciplinary team consensus Authors: Jingjing Yu, Peijun Hu, Ruixia Dai, Xiaomin Liu, Shanshan Zhang, Yangdi Xu, Yiwei Gao, Bingjian Lu, Wenqian Wang, Yizhen Niu, Xiaochen Wang, Shihao Xu, Weibin Wang, Qianqian Qi, Fan Liu, Yang Zhang, Sihui Hu, Yu Tian, Weiguo Lu, Jingsong Li, Jiale Qin Published: 2026 ABSTRACT Ovarian cancer, with high mortality, demands accurate preoperative assessment to guide individualized treatment. It typically requires ultrasound, CT, and MRI interpreted by multidisciplinary teams (MDTs) in complex cases. We developed OVUCM, a multi-task AI system that integrates multi-modalities via intermediate fusion using radiomics, machine learning, and 5×4 nested cross-validation. Trained on 1742 patients from a cancer center, OVUCM predicted five clinical tasks with AUCs of 0.847–0.929: benign vs. non-benign, borderline vs. malignant, non-epithelial vs. epithelial, FIGO stages I-II vs. III-IV, and non-HGSOC vs. HGSOC. External validation in 150 patients from two general hospitals confirmed generalizability (AUCs: 0.833–0.974). The system achieved diagnostic parity with MDT consensus in four tasks and outperformed it in one, while consistently surpassing at least one independent gynecologist across all tasks. By emulating MDT-level interpretation, OVUCM bridges the gap between single-modality tools and comprehensive clinical decision-making, offering a scalable solution in resource-limited settings. Introduction Ovarian cancer remains the leading cause of mortality among malignant tumors of the female reproductive system, with a five-year survival rate of only 30–50%, posing a severe threat to women’s health1. Its therapeutic advances now enable personalized management based on tumor biology, including the nature of the mass (borderline or malignant), pathological classification (epithelial or non-epithelial, high-grade serous ovarian cancer (HGSOC) or non-HGSOC), and International Federation of Gynecology and Obstetrics (FIGO) staging2–5. However, in current clinical practice, these individualized factors are predominantly obtained through intraoperative clinical-pathological staging and postoperative tissue analysis4–7. Notably, preoperative fine-needle aspiration biopsy is contraindicated in cases of early-stage cancer and cystic ovarian masses. Clinical guidelines from the National Comprehensive Cancer Network6, the European Society for Medical Oncology8, and the Japan Society of Gynecologic Oncology9 advise against invasive procedures due to the risk of tumor capsule rupture and peritoneal dissemination of malignant cells, which can artificially elevate clinical staging and adversely impact prognosis. In light of these challenges, utilizing the non-invasive capabilities of imaging modalities to acquire critical preoperative information for personalized treatment may constitute an effective strategy for optimizing therapeutic outcomes. In routine practice, the pre-treatment imaging assessment of ovarian cancer typically involves pelvic ultrasound, computed tomography (CT), or magnetic resonance imaging (MRI), with the choice guided by the clinical question and institutional protocols6. Ultrasound is generally the first-line modality for detecting adnexal masses and distinguishing between benign and malignant lesions, with color Doppler providing valuable insights into tumor neovascularization. In specialized centers, ultrasound is also used to assess disease extent10–13. CT, especially contrast-enhanced CT (CECT), is predominantly used for identifying peritoneal metastases and preoperative clinical staging, although it tends to underestimate the stages by approximately 20.9%13,14. MRI, as second-line imaging, is recommended for evaluating the indeterminate lesions on ultrasound, and is currently regraded as the most accurate imaging modality for assessing mesenteric root infiltration and subtle peritoneal metastases15–19. Together, these imaging modalities provide anatomical, functional, and dynamic information about ovarian masses, potentially enhancing diagnostic accuracy and informing individualized treatment planning6. However, integrating imaging data into clinical decision-making remains complex and primarily relies on the multidisciplinary team (MDT), comprising gynecological oncologists, ultrasound physicians, and radiologists20,21. The variability in diagnostic accuracy is influenced by factors such as the institutional status as a specialized cancer center, the expertise of physicians, and the inherent subjectivity of image interpretation. These challenges highlight the unmet need for more objective and quantitative integration of multi-modality imaging techniques to improve preoperative assessment and guide personalized treatment strategies. Radiomics enables the extraction of extensive standardized biomarker data from medical images22,23, and when paired with artificial intelligence (AI) techniques, like deep learning and machine learning, it facilitates high-throughput analysis of image features that reflect both phenotypic and, potentially, molecular or genetic tumor heterogeneity23–26. These advancements have accelerated new forms of data-driven insights for precise tumor diagnosis. Most studies to date have demonstrated the efficacy of these techniques in extracting diagnostic information and addressing the heterogeneity of ovarian cancer. For instance, ultrasound radiomics with deep learning and machine learning classified benign and malignant ovarian masses27–32. Similarly, CT radiomics with machine learning distinguished between benign and malignant ovarian tumors33. Additionally, MRI radiomics, in combination with machine learning, has revealed characteristics of histological subtypes, such as HGSOC34. However, most approaches are restricted to single-modality imaging data and are typically designed to address a single clinical task, predominantly focusing on distinguishing malignant from benign lesions35–37. Integrated approaches combining multiple modalities remain underdeveloped, yet there is a critical need for a preoperative fusion method for ovarian cancer that integrates ultrasound, CT, and MRI. This approach should align with clinical workflows, providing comprehensive diagnostic insights into tumor nature, histological subtypes, and FIGO staging to enhance diagnostic precision and support individualized treatment planning. In this study, we developed an OVUCM system, referred to as a multi-task preoperative prediction system for OVarian cancer through integrating Ultrasound, CECT, and MRI images, alongside clinical text information for patients from Women’s Hospital, Zhejiang University School of Medicine. This hospital is the National Regional Medical Center for Obstetrics and Gynecology, and the National Key Specialty in Gynecological Oncology (referred to as the Cancer Center). The integration framework employed radiomics for imaging feature extraction, machine learning classifiers, and an intermediate fusion approach (Fig. 1). The system sequentially predicted multiple outcomes, including benign vs. non-benign, borderline vs. malignant, non-epithelial vs. epithelial, FIGO staging I-II vs. FIGO III-IV, and non-HGSOC vs. HGSOC. Its performance was compared to that of MDT and individual gynecologists at the Cancer Center, while its external generalization capability was evaluated in two general hospitals (non-specialized cancer center). Our findings empirically demonstrated the advantages of the OVUCM system in enhancing preoperative stratification and supporting personalized treatment decision-making in clinical practice.Fig. 1Schematic outline of the study design.The figure illustrates the development workflow of the OVUCM system. Each patient underwent ultrasound, CECT, T2WI, and DWI examinations. From grayscale and color Doppler ultrasound, 464 radiomic features were extracted per modality, while 1,288 features were derived from each of the three cross-sectional imaging modalities (CECT, T2WI, and DWI). A four-step feature selection pipeline was applied to reduce redundancy and rank feature importance, resulting in filtered feature pools for every single modality. Then, top k ranked features (k in [5, 10, 15, 20, 30, 40, 50]) along with five machine learning classifiers (LinR, LogR, SVC, NB, XGB) were used to construct single-modality candidate models. Using a 5×4 nested cross-validation, the optimal single-modality model with the feature number, classifier, and classifier hyperparameters was determined with respect to the highest AUC value. The corresponding features from these optimal models formed the multi-modality feature pool for 11 modality combinations. Subsequently, multi-modality intermediate fusions, including dual-, triple-, and quadruple-modality combinations, were constructed in steps. For each combination, the same modeling strategy was applied: top-k ranked features were selected from modality-specific pools, and candidate models were trained using the five classifiers with 5×4 nested cross-validation. For each of the five clinical tasks, the optimal fusion models achieving the highest AUC were selected. Finally, clinical characteristics (e.g., age, menopausal status) and serum tumor markers (e.g., CA125) were integrated into the five task-specific fusion models to develop the OVUCM system. Created using BioRender.com. AFP, alpha-fetoprotein; BMI, body mass index; CEA, carcinoembryonic antigen; CECT, contrast-enhanced computed tomography; DWI, MRI diffusion-weighted imaging; FIGO, International Federation of Gynecology and Obstetrics; GLCM, gray level co-occurrence matrix; GLRLM, gray level run length matrix; HGSOC, high-grade serous ovarian cancer; LinR, Linear Regression; Log R, Logistic Regression; MDT, multidisciplinary team; MRI, magnetic resonance imaging; NB, Naive Bayes; SCC, squamous cell carcinoma antigen; SVC, Support Vector Classifier; T2WI, MRI-T2-weighted imaging; XGB, eXtreme Gradient Boost. Schematic outline of the study design. The figure illustrates the development workflow of the OVUCM system. Each patient underwent ultrasound, CECT, T2WI, and DWI examinations. From grayscale and color Doppler ultrasound, 464 radiomic features were extracted per modality, while 1,288 features were derived from each of the three cross-sectional imaging modalities (CECT, T2WI, and DWI). A four-step feature selection pipeline was applied to reduce redundancy and rank feature importance, resulting in filtered feature pools for every single modality. Then, top k ranked features (k in [5, 10, 15, 20, 30, 40, 50]) along with five machine learning classifiers (LinR, LogR, SVC, NB, XGB) were used to construct single-modality candidate models. Using a 5×4 nested cross-validation, the optimal single-modality model with the feature number, classifier, and classifier hyperparameters was determined with respect to the highest AUC value. The corresponding features from these optimal models formed the multi-modality feature pool for 11 modality combinations. Subsequently, multi-modality intermediate fusions, including dual-, triple-, and quadruple-modality combinations, were constructed in steps. For each combination, the same modeling strategy was applied: top-k ranked features were selected from modality-specific pools, and candidate models were trained using the five classifiers with 5×4 nested cross-validation. For each of the five clinical tasks, the optimal fusion models achieving the highest AUC were selected. Finally, clinical characteristics (e.g., age, menopausal status) and serum tumor markers (e.g., CA125) were integrated into the five task-specific fusion models to develop the OVUCM system. Created using BioRender.com. AFP, alpha-fetoprotein; BMI, body mass index; CEA, carcinoembryonic antigen; CECT, contrast-enhanced computed tomography; DWI, MRI diffusion-weighted imaging; FIGO, International Federation of Gynecology and Obstetrics; GLCM, gray level co-occurrence matrix; GLRLM, gray level run length matrix; HGSOC, high-grade serous ovarian cancer; LinR, Linear Regression; Log R, Logistic Regression; MDT, multidisciplinary team; MRI, magnetic resonance imaging; NB, Naive Bayes; SCC, squamous cell carcinoma antigen; SVC, Support Vector Classifier; T2WI, MRI-T2-weighted imaging; XGB, eXtreme Gradient Boost. Results Cohort and clinical characteristics The study included 1,892 female patients with ovarian mass, with a median age at diagnosis of 48 years (interquartile range [IQR] 33-57). Histological analysis confirmed 1,125 benign (59.46%), 197 borderline (10.41%), and 570 malignant (30.13%) lesions, with detailed histological diagnosis provided in Supplementary Table 1. Among them, 1,742 patients were recruited from the Cancer Center and were used for model training, validation, and internal testing. An additional 102 patients were sourced from The Fourth Affiliated Hospital, Zhejiang University School of Medicine (ZJU4H), and 48 cases were collected from The First Affiliated Hospital of Wenzhou Medical University (FHWMU). Both are general hospitals and served as the external test sets 1 and 2, respectively. The clinical characteristics for three cohorts are detailed in Table 1. A flowchart depicting patient inclusion is illustrated in Fig. 2.Fig. 2Flowchart of patient inclusion and exclusion criteria.A total of 1742 patients from internal datasets (WHZJU), who underwent preoperative ultrasound, CECT, MRI, were used to develop a multi-task OVUCM system. Each task was tested using both internal and external datasets. The independent external patients (ZJU4H and FHWMU) were included in each task based on the specific combination of modalities required. For example, the model for Task 1 included ultrasound, CECT and MRI, while Task 2 included ultrasound and MRI. FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer; WHZJU, Women’s Hospital, Zhejiang University School of Medicine; ZJU4H, the Fourth Affiliated Hospital, Zhejiang University School of Medicine; FHWMU, the First Affiliated Hospital of Wenzhou Medical University.Table 1Clinical characteristics of patients in the internal and external datasetsVariablesInternal datasets (n = 1742)External datasets 1 (n = 102)External datasets 2 (n = 48)Age (years)47 (33–57)52 (39–57)58 (46–62)BMI (kg/m2)22.59 (20.83–24.77)23.44 (21.63–26.29)23.68 (21.23–25.81)Menopausal status Postmenopausal706 (40.53%)56 (54.90%)37 (77.08%) Premenopausal1036 (59.47%)46 (45.10%)11 (22.92%)Gravida2 (1–3)2 (2–4)3 (2–4) 0352 (20.21%)10 (9.80%)4 (8.33%) 1263 (15.10%)13 (12.75%)3 (6.25%) ≥21,127 (64.70%)79 (77.45%)41 (85.42%)Parity1 (1–2)2 (1–2)2 (1–3) 0408 (23.42%)17 (16.67%)4 (8.33%) 1832 (47.76%)27 (26.47%)10 (20.83%) ≥2502 (28.82%)58 (56.86%)34 (70.83%)CA12525.10 (13.20–83.10)38.12 (16.36–156.89)78.7 (18.00–317.90) ≤25.0 U/mL772 (44.32%)34 (33.33%)15 (31.25%) >25.0 U/mL825 (47.36%)53 (51.96%)31 (64.58%) N.A.145 (8.32%)15 (14.71%)2 (4.17%)CA19911.00 (6.60–22.48)7.99 (3.35–28.60)9.50 (4.20–23.00) ≤30.0 U/mL1,306 (74.97%)65 (63.73%)34 (70.83%) >30.0 U/mL263 (15.10%)21 (20.59%)10 (20.83%) N.A.173 (9.93%)16 (15.69%)4 (8.33%)AFP2.30 (1.60–3.30)2.19 (1.73–3.08%)2.75 (1.87–3.81) ≤7.0 ng/mL1,504 (86.34%)82 (80.39%)44 (91.67%) >7.0 ng/mL45 (2.58%)4 (3.92%)2 (4.17%) N.A.193 (11.08%)16 (15.69%)2 (4.17%)CEA1.30 (0.80–2.00)1.65 (1.20–2.24)1.60 (0.90–3.00) ≤5.0 ng/mL1,514 (86.91%)82 (80.39%)37 (77.08%) >5.0 ng/mL43 (2.47%)4 (3.92%)8 (16.67%) N.A.185 (10.62%)16 (15.69%)3 (6.25%)CA15310.40 (7.10–16.88)11.00 (6.99–18.25)11.70 (7.10–22.90) ≤24.0 U/mL1299 (74.57%)69 (67.65%)27 (56.25%) >24.0 U/mL213 (12.23%)16 (15.69%)14 (29.17%) N.A.230 (13.20%)17 (16.67%)7 (14.58%)SCC0.90 (0.70–1.30)0.83 (0.60–1.10)0.80 (0.60–1.20) <2.7 ng/mL1468 (84.27%)83 (81.37%)39 (81.25%) ≥2.7 ng/mL39 (2.24%)3 (2.94%)1 (2.08%) N.A.235 (13.49%)16 (15.69%)8 (16.67%)Histological diagnoses Benign1064 (61.08%)45 (44.12%)16 (33.33%) Borderline173 (9.93%)19 (18.63%)5 (10.42%) Malignant505 (28.99%)38 (37.25%)27 (56.25%)Histological subtypes Epithelial tumors934 (53.62%)64 (62.75%)34 (70.83%) Serous tumors551 (31.63%)46 (45.10%)28 (58.33%) Mucinous tumors210 (12.06%)13 (12.75%)4 (8.33%) Endometrioid tumors43 (2.47%)2 (1.96%)0 (0.00%) Clear cell tumors55 (3.16%)1 (0.98%)1 (2.08%) Othera75 (4.30%)2 (1.96%)1 (2.08%) Non-epithelial tumors808 (46.39%)38 (37.24%)14 (29.17%) Sex cord-stromal tumors189 (10.85%)12 (11.76%)5 (10.42%) Germ cell tumors210 (12.05%)7 (6.86%)4 (8.33%) Tumor-like lesions229 (13.15%)10 (9.80%)4 (8.33%) Ovarian endometriotic cysts171 (9.82%)9 (8.82%)1 (2.08%) Otherb9 (0.52%)0 (0.00%)0 (0.00%)Continuous data are given as median (IQR) and categorical data as n (%).AFP alpha-fetoprotein, BMI body mass index, CEA carcinoembryonic antigen, SCC squamous cell carcinoma antigen.a Other epithelial tumor types include seromucinous tumors, Brenner tumors, and others.b Other non-epithelial tumor types include mesothelioma, miscellaneous, mesenchymal, and others. The detailed histological diagnoses are provided in Supplementary Table 1. Flowchart of patient inclusion and exclusion criteria. A total of 1742 patients from internal datasets (WHZJU), who underwent preoperative ultrasound, CECT, MRI, were used to develop a multi-task OVUCM system. Each task was tested using both internal and external datasets. The independent external patients (ZJU4H and FHWMU) were included in each task based on the specific combination of modalities required. For example, the model for Task 1 included ultrasound, CECT and MRI, while Task 2 included ultrasound and MRI. FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer; WHZJU, Women’s Hospital, Zhejiang University School of Medicine; ZJU4H, the Fourth Affiliated Hospital, Zhejiang University School of Medicine; FHWMU, the First Affiliated Hospital of Wenzhou Medical University. Table 1Clinical characteristics of patients in the internal and external datasetsVariablesInternal datasets (n = 1742)External datasets 1 (n = 102)External datasets 2 (n = 48)Age (years)47 (33–57)52 (39–57)58 (46–62)BMI (kg/m2)22.59 (20.83–24.77)23.44 (21.63–26.29)23.68 (21.23–25.81)Menopausal status Postmenopausal706 (40.53%)56 (54.90%)37 (77.08%) Premenopausal1036 (59.47%)46 (45.10%)11 (22.92%)Gravida2 (1–3)2 (2–4)3 (2–4) 0352 (20.21%)10 (9.80%)4 (8.33%) 1263 (15.10%)13 (12.75%)3 (6.25%) ≥21,127 (64.70%)79 (77.45%)41 (85.42%)Parity1 (1–2)2 (1–2)2 (1–3) 0408 (23.42%)17 (16.67%)4 (8.33%) 1832 (47.76%)27 (26.47%)10 (20.83%) ≥2502 (28.82%)58 (56.86%)34 (70.83%)CA12525.10 (13.20–83.10)38.12 (16.36–156.89)78.7 (18.00–317.90) ≤25.0 U/mL772 (44.32%)34 (33.33%)15 (31.25%) >25.0 U/mL825 (47.36%)53 (51.96%)31 (64.58%) N.A.145 (8.32%)15 (14.71%)2 (4.17%)CA19911.00 (6.60–22.48)7.99 (3.35–28.60)9.50 (4.20–23.00) ≤30.0 U/mL1,306 (74.97%)65 (63.73%)34 (70.83%) >30.0 U/mL263 (15.10%)21 (20.59%)10 (20.83%) N.A.173 (9.93%)16 (15.69%)4 (8.33%)AFP2.30 (1.60–3.30)2.19 (1.73–3.08%)2.75 (1.87–3.81) ≤7.0 ng/mL1,504 (86.34%)82 (80.39%)44 (91.67%) >7.0 ng/mL45 (2.58%)4 (3.92%)2 (4.17%) N.A.193 (11.08%)16 (15.69%)2 (4.17%)CEA1.30 (0.80–2.00)1.65 (1.20–2.24)1.60 (0.90–3.00) ≤5.0 ng/mL1,514 (86.91%)82 (80.39%)37 (77.08%) >5.0 ng/mL43 (2.47%)4 (3.92%)8 (16.67%) N.A.185 (10.62%)16 (15.69%)3 (6.25%)CA15310.40 (7.10–16.88)11.00 (6.99–18.25)11.70 (7.10–22.90) ≤24.0 U/mL1299 (74.57%)69 (67.65%)27 (56.25%) >24.0 U/mL213 (12.23%)16 (15.69%)14 (29.17%) N.A.230 (13.20%)17 (16.67%)7 (14.58%)SCC0.90 (0.70–1.30)0.83 (0.60–1.10)0.80 (0.60–1.20) <2.7 ng/mL1468 (84.27%)83 (81.37%)39 (81.25%) ≥2.7 ng/mL39 (2.24%)3 (2.94%)1 (2.08%) N.A.235 (13.49%)16 (15.69%)8 (16.67%)Histological diagnoses Benign1064 (61.08%)45 (44.12%)16 (33.33%) Borderline173 (9.93%)19 (18.63%)5 (10.42%) Malignant505 (28.99%)38 (37.25%)27 (56.25%)Histological subtypes Epithelial tumors934 (53.62%)64 (62.75%)34 (70.83%) Serous tumors551 (31.63%)46 (45.10%)28 (58.33%) Mucinous tumors210 (12.06%)13 (12.75%)4 (8.33%) Endometrioid tumors43 (2.47%)2 (1.96%)0 (0.00%) Clear cell tumors55 (3.16%)1 (0.98%)1 (2.08%) Othera75 (4.30%)2 (1.96%)1 (2.08%) Non-epithelial tumors808 (46.39%)38 (37.24%)14 (29.17%) Sex cord-stromal tumors189 (10.85%)12 (11.76%)5 (10.42%) Germ cell tumors210 (12.05%)7 (6.86%)4 (8.33%) Tumor-like lesions229 (13.15%)10 (9.80%)4 (8.33%) Ovarian endometriotic cysts171 (9.82%)9 (8.82%)1 (2.08%) Otherb9 (0.52%)0 (0.00%)0 (0.00%)Continuous data are given as median (IQR) and categorical data as n (%).AFP alpha-fetoprotein, BMI body mass index, CEA carcinoembryonic antigen, SCC squamous cell carcinoma antigen.a Other epithelial tumor types include seromucinous tumors, Brenner tumors, and others.b Other non-epithelial tumor types include mesothelioma, miscellaneous, mesenchymal, and others. The detailed histological diagnoses are provided in Supplementary Table 1. Clinical characteristics of patients in the internal and external datasets Continuous data are given as median (IQR) and categorical data as n (%). AFP alpha-fetoprotein, BMI body mass index, CEA carcinoembryonic antigen, SCC squamous cell carcinoma antigen. a Other epithelial tumor types include seromucinous tumors, Brenner tumors, and others. b Other non-epithelial tumor types include mesothelioma, miscellaneous, mesenchymal, and others. The detailed histological diagnoses are provided in Supplementary Table 1. Radiomics feature extraction and selection The ultrasound dataset comprised 5535 grayscale and 4777 color Doppler images, with an average of 2.9 grayscale (range: 1–32) and 2.5 color Doppler images (range: 0–22) per patient. Each patient also underwent one CECT and one MRI examination (n = 1892 each), yielding an average of 85 axial CT slices and 48 MRI slices (including T2WI and DWI sequences). For the 149 cases lacking color Doppler images, median imputation was performed using values from cases with the same labels. Ultimately, each patient was characterized by 4792 radiomic features: 464 features extracted from grayscale images, 464 from color Doppler, and 1288 each from CECT, T2-weighted MRI (T2WI), and diffusion-weighted imaging MRI (DWI) images. CT and MRI images were segmented using automated models, which achieved robust performance with Dice coefficients of 0.863 and 0.846, respectively. Ultrasound-CECT-MRI fusion and OVUCM system development We first constructed single-modality models using US, CECT, T2WI, and DWI data across five clinical tasks. For each task, the optimal single-modality mode was selected based on the highest area under the curve (AUC). For example, in Task 1 (benign vs. non-benign masses), the CECT-based model using the XGB classifier achieved the highest AUC of 0.876 (Supplementary Table 2), with 50 radiomic features selected and carried forward for further analysis. Subsequently, 11 multi-modality fusion strategies were evaluated. Consistent with the single-modality approach, the optimal fusion of imaging modalities and classifier was selected based on the highest AUC. For instance, in Task 1, the fusion of ultrasound, CECT, and DWI under the XGB classifier achieved the highest AUC of 0.916. The optimal multi-modality models for the remaining four tasks achieved the highest AUCs, ranging from 0.804 to 0.846 (Fig. 3 and Supplementary Table 2). These five task-specific fusion models were then integrated with clinical variables to construct the OVUCM system.Fig. 3Visualization of the importance of modalities and features.The Histograms display the top 10 most important features for each task, calculated using the Permutation Importance algorithm. The vertical ordinate of histograms shows the feature names in order of importance, while the horizontal one shows the feature importance score. Pie charts show the proportion of image modalities contributed to each task. CECT, contrast-enhanced computed tomography; DWI, MRI diffusion-weighted imaging; FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer; T2WI, MRI-T2-weighted imaging; USG, ultrasound grayscale imaging; USC, ultrasound color Doppler imaging. Visualization of the importance of modalities and features. The Histograms display the top 10 most important features for each task, calculated using the Permutation Importance algorithm. The vertical ordinate of histograms shows the feature names in order of importance, while the horizontal one shows the feature importance score. Pie charts show the proportion of image modalities contributed to each task. CECT, contrast-enhanced computed tomography; DWI, MRI diffusion-weighted imaging; FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer; T2WI, MRI-T2-weighted imaging; USG, ultrasound grayscale imaging; USC, ultrasound color Doppler imaging. OVUCM system evaluation In internal datasets, the OVUCM system demonstrated consistently favorable performance across five clinical tasks, with 5-fold mean AUCs of 0.929 (95% CI: 0.918–0.939) for benign vs. non-benign, 0.889 (0.872–0.909) for borderline vs. malignant, 0.897 (0.863–0.934) for non-epithelial vs. epithelial, 0.853 (0.841–0.862) for FIGO staging I-II vs. III-IV, and 0.847 (0.808–0.890) for non-HGSOC vs. HGSOC (Fig. 4a and Table 2). Supplementary Table 3 presents detailed metrics, including cutoff value, accuracy, sensitivity, specificity, precision, and F1-score across cross-validation folds. Confusion matrix revealed high classification consistency, particularly for true negative for Task 1 and true positives in Tasks 2–5 (Fig. 4b). Calibration curves indicated a close alignment between predicted probabilities and actual outcomes across five clinical tasks, with Brier scores ranging from 0.092 to 0.216 (Supplementary Fig. 1). Decision Curve Analysis (DCA) showed that OVUCM-guided decisions provided greater net clinical benefits than ‘treat-all’ and ‘treat-none’ approaches ranges (0.07–0.99 across tasks, Supplementary Fig. 2).Fig. 4Classification performance for the OVUCM system on both internal and external datasets.The ROC curves of the internal datasets (a) show 5-fold AUCs of 5 × 4 nested cross-validation along with their mean values, while the external datasets 1 (c) and the external datasets 2 (e) present the tested performance on two independent cohorts. The confusion matrices offer a detailed breakdown of classification outcomes, including true positives, false positives, true negatives, and false negatives, for providing deeper insights into performance interpretation in internal (b) and external (d, f) datasets. AUC, area under the curve; FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer; ROC, receiver operating characteristic curve.Table 2Selected modalities by OVUCM system and their performance in each task across internal and external datasetsTask 1(Benign vs. Non-benign)Task 2(Borderline vs. Malignant)Task 3(Non-epithelial vs. Epithelial)Task 4(FIGO I-II vs. FIGO III-IV)Task 5(Non-HGSOC vs. HGSOC)InternalExternal 1External 2InternalExternal 1External 2InternalExternal 1External 2InternalExternal 1External 2InternalExternal 1External 2ModalityClinical, US, CECT, DWIClinical, US, DWIClinical, US, CECT, DWIClinical, CECT, DWI, T2WIClinical, T2WICutoff0.3130.4630.6070.4670.347AUC (95% CI)0.929 (0.918–0.939)0.899 (0.828–0.956)0.931 (0.823–1.000)0.889 (0.872–0.909)0.935 (0.860–0.986)0.833 (0.513–1.000)0.897 (0.863–0.934)0.918 (0.810–1.000)0.867 (0.625–1.000)0.853 (0.841–0.862)0.974 (0.905–1.000)0.864 (0.633–1.000)0.847 (0.808–0.890)0.969 (0.867–1.000)0.929 (0.769–1.000)Accuracy0.8520.8120.8820.8350.8520.9310.8710.8390.8240.8060.8850.8820.8200.9440.938Sensitivity0.8840.8250.8330.8280.8570.9580.8780.8260.8000.8190.8820.9090.8190.9380.929Specificity0.8320.8000.9380.8550.8420.8000.8320.8751.0000.7920.8890.8330.8241.0001.000Precision0.7710.7860.9380.9440.9090.9580.9700.9501.0000.8280.9380.9090.9661.0001.000F1–score0.8230.8050.8820.8800.8820.9580.9200.8840.8890.8170.9090.9090.8810.9680.963AUC area under the curve, CECT contrast-enhanced computed tomography, DWI MRI diffusion-weighted imaging, FIGO International Federation of Gynecology and Obstetrics, HGSOC high-grade serous ovarian cancer, T2WI MRI-T2-weighted imaging, US ultrasound. Classification performance for the OVUCM system on both internal and external datasets. The ROC curves of the internal datasets (a) show 5-fold AUCs of 5 × 4 nested cross-validation along with their mean values, while the external datasets 1 (c) and the external datasets 2 (e) present the tested performance on two independent cohorts. The confusion matrices offer a detailed breakdown of classification outcomes, including true positives, false positives, true negatives, and false negatives, for providing deeper insights into performance interpretation in internal (b) and external (d, f) datasets. AUC, area under the curve; FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer; ROC, receiver operating characteristic curve. Table 2Selected modalities by OVUCM system and their performance in each task across internal and external datasetsTask 1(Benign vs. Non-benign)Task 2(Borderline vs. Malignant)Task 3(Non-epithelial vs. Epithelial)Task 4(FIGO I-II vs. FIGO III-IV)Task 5(Non-HGSOC vs. HGSOC)InternalExternal 1External 2InternalExternal 1External 2InternalExternal 1External 2InternalExternal 1External 2InternalExternal 1External 2ModalityClinical, US, CECT, DWIClinical, US, DWIClinical, US, CECT, DWIClinical, CECT, DWI, T2WIClinical, T2WICutoff0.3130.4630.6070.4670.347AUC (95% CI)0.929 (0.918–0.939)0.899 (0.828–0.956)0.931 (0.823–1.000)0.889 (0.872–0.909)0.935 (0.860–0.986)0.833 (0.513–1.000)0.897 (0.863–0.934)0.918 (0.810–1.000)0.867 (0.625–1.000)0.853 (0.841–0.862)0.974 (0.905–1.000)0.864 (0.633–1.000)0.847 (0.808–0.890)0.969 (0.867–1.000)0.929 (0.769–1.000)Accuracy0.8520.8120.8820.8350.8520.9310.8710.8390.8240.8060.8850.8820.8200.9440.938Sensitivity0.8840.8250.8330.8280.8570.9580.8780.8260.8000.8190.8820.9090.8190.9380.929Specificity0.8320.8000.9380.8550.8420.8000.8320.8751.0000.7920.8890.8330.8241.0001.000Precision0.7710.7860.9380.9440.9090.9580.9700.9501.0000.8280.9380.9090.9661.0001.000F1–score0.8230.8050.8820.8800.8820.9580.9200.8840.8890.8170.9090.9090.8810.9680.963AUC area under the curve, CECT contrast-enhanced computed tomography, DWI MRI diffusion-weighted imaging, FIGO International Federation of Gynecology and Obstetrics, HGSOC high-grade serous ovarian cancer, T2WI MRI-T2-weighted imaging, US ultrasound. Selected modalities by OVUCM system and their performance in each task across internal and external datasets AUC area under the curve, CECT contrast-enhanced computed tomography, DWI MRI diffusion-weighted imaging, FIGO International Federation of Gynecology and Obstetrics, HGSOC high-grade serous ovarian cancer, T2WI MRI-T2-weighted imaging, US ultrasound. To emphasize real-world applicability, we focused in detail on the model’s performance in two external cohorts from general hospitals (ZJU4H and FHWMU). The system maintained strong discriminative performance across all tasks, with AUC ranges of: Task 1 (0.899–0.931), Task 2 (0.833–0.935), Task 3 (0.867–0.918), Task 4 (0.864–0.974), and Task 5 (0.929–0.969) (Fig. 4c, e and Table 2). Other metrics (e.g., sensitivity and F1-score) were comparable to those of internal validation. The confusion matrices confirmed consistent classification trends with internal data. Given the relatively small sample sizes and high positive case ratios in the two external sets, we combined them to enhance the statistical reliability of the calibration curves and DCA, and to improve the evaluation of the model’s generalizability. The integrated external calibration curves demonstrated that the system had Brier scores below 0.25 across all five tasks (range from 0.136 to 0.210; Supplementary Fig. 3). DCA showed net clinical benefit over default approaches within meaningful threshold ranges: 0.20–0.97, 0.04–0.96, 0.20–0.93, 0.16–0.99, and 0.51–0.99 across the five Tasks (Supplementary Fig. 4). These findings support that the OVUCM system, trained on Cancer Center data, may be generalized to non-specialized clinical settings not involved in model development. Comparison of the OVUCM system with clinicians The comparison study was conducted on 35 paired patients, with their hierarchical distribution shown in Supplementary Fig. 5 and clinical characteristics listed in Supplementary Table 4. When compared to MDT from the Cancer Center, the OVUCM system showed no statistically significant differences in the pairwise AUCs for tasks 1-4 (ΔAUC 0.025–0.300, P > 0.05), and the higher AUC in Task 5 (P < 0.05). It indicated that the OVUCM system demonstrated comparable diagnostic performance with MDT. In comparison with three individual gynecologists, the AUC values of the OVUCM system in all tasks significantly outperformed those of one or more clinicians (Fig. 5). Notably, for histological subtypes discrimination, the system outperformed all clinicians by a large margin: in Task 3 (non-epithelial vs. epithelial), 0.525–0.725 vs. 1.000; and in Task 5 (non-HGSOC vs. HGSOC), 0.450–0.550 vs. 1.000, respectively (all P < 0.05, Supplementary Table 5).Fig. 5The ROC curves of comparative study between the OVUCM system, MDT and individual gynecologists.OVUCM, a multitask system for OVarian cancer through integration of Ultrasound, CECT, and MRI imaging; AUC, area under the curve; FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer; MDT, Multidisciplinary Team. The ROC curves of comparative study between the OVUCM system, MDT and individual gynecologists. OVUCM, a multitask system for OVarian cancer through integration of Ultrasound, CECT, and MRI imaging; AUC, area under the curve; FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer; MDT, Multidisciplinary Team. The nightingale rose diagrams (Fig. 6) and radar charts (Supplementary Fig. 6) further visually illustrated these findings. Across key metrics including AUC, sensitivity, and specificity, the OVUCM system closely approximated MDT performance and consistently outperformed individual gynecologists in several tasks. The radar charts further highlighted the OVUCM system’s robustness across increasing diagnostic complexity. While gynecologists performed well in detecting non-benign masses, their performance declined notably as clinical stratification increased.Fig. 6Comparative study between the OVUCM system, MDT and individual gynecologists.The nightingale rose diagrams, depicting the AUC (a), sensitivity (b), and specificity (c) values of the OVUCM system, MDT, and three individual gynecologists. OVUCM, a multitask system for OVarian cancer through integration of Ultrasound, CECT, and MRI imaging; AUC, area under the curve; FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer; MDT, Multidisciplinary Team. Comparative study between the OVUCM system, MDT and individual gynecologists. The nightingale rose diagrams, depicting the AUC (a), sensitivity (b), and specificity (c) values of the OVUCM system, MDT, and three individual gynecologists. OVUCM, a multitask system for OVarian cancer through integration of Ultrasound, CECT, and MRI imaging; AUC, area under the curve; FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer; MDT, Multidisciplinary Team. Discussion The development of a multimodal, multi-task ovarian cancer diagnostic system represents a critical step toward precision oncology in ovarian cancer. However, at present, even when these data are available, they are rarely integrated into a unified diagnostic framework. In this study, we presented the OVUCM system, integrating ultrasound-CT-MRI data available in the routine practice. This system showed favorable diagnostic performance across five tasks (AUC: 0.928–1.000), approaching Cancer Center MDT evaluation. External validation in non-specialized hospitals further indicated potential robustness across different clinical environments. By unifying intermediate fusion, nested cross-validation, and task-aware modality prioritization, OVUCM provides a proof-of-concept framework for multimodal integration in preoperative assessment and shows promise as a clinical AI tool. In clinical practice, preoperative ultrasound, CT, and MRI each provide complementary disease features of ovarian cancer evaluation. Integrating these features holds significant clinical value, but the challenge lies in effectively fusing them to maximize their potential. To address this, we developed a unified framework of image data annotation, feature extraction and selection, and multimodal fusion modeling. The intermediate fusion strategy used in this framework strikes a balance between modality-specific feature preservation (mitigating information loss in early fusion) and model robustness (reducing overfitting risks inherent in late fusion)38,39. Additionally, we utilized machine learning algorithms to model these multimodal features, and nested cross-validation for flexible training and optimal model across different tasks, as well as feature importance analysis. Notably, the OVUCM system dynamically prioritizes imaging modalities according to clinical tasks, emulating human diagnostic reasoning. For example, in FIGO staging (Task 4), CT and MRI took precedence over ultrasound, aligning with their established role in detecting peritoneal metastases and lymph node involvement40,41. Moreover, CECT-derived tumor sphericity was identified as the dominant feature, a biomarker previously associated with advanced disease and chemoresistance42–44. The system’s interpretable modality and feature selection bridge radiomics with tumor biology, addressing the “black box” limitations of conventional AI. Ovarian cancer stratification remains a key area of ongoing research. Ultrasound-based tools such as IOTA-ADNEX and O-RADS US represent landmark advances in the standardized characterization and risk stratification of adnexal masses, as highlighted by major prospective studies45–47. In comparison, we conducted a head-to-head retrospective evaluation of the OVUCM system against clinicians applying these established models in the same cohort. The OVUCM system demonstrated sensitivity comparable to IOTA-ADNEX and O-RADS US at the 10% threshold (all P > 0.05). Regarding specificity, our ultrasound physicians showed numerically lower values than those reported in the prospective screening studies (0.467–0.578 vs. 0.585–0.783). This difference is likely explained by the higher prevalence of non-benign lesions in the cohort (47.06% vs 10.04–32.65%), and a greater proportion of benign tumors with atypical or complex imaging features that often require further evaluation (CT/MRI), which inherently lowers specificity48. Beyond initial differentiation, OVUCM also appears to offer advantages in more complex prognostic tasks. For the challenging task of distinguishing borderline from malignant tumors, OVUCM achieved an AUC of 0.833–0.935 and sensitivity of 0.857–0.958, which tends to be higher than previously reported models (AUC 0.71–0.91) and AI-based approaches (sensitivity: 68.35–69.38%)27,47. In FIGO staging, OVUCM’s integration of multimodal imaging features enabled an early- vs. advanced-stage classification accuracy of 88.2–88.5%, exceeding the reported accuracy of standalone CT or MRI (accuracy: 77-78%)49. The system further demonstrated encouraging results in histological subtype prediction (AUC = 0.867–0.918; Specificity = 0.875–1.000) for differentiating epithelial from non-epithelial tumors. While its performance in individual tasks was encouraging, the primary strength of the OVUCM system lies in its unified, multi-task framework that supports simultaneous preoperative prediction of tumor nature (benign/borderline/malignant), FIGO staging, and histological subtypes. This multi-faceted capability suggests that the OVUCM system could serve as a valuable framework in ovarian cancer precise diagnostics. In current clinical practice, optimal diagnosis of ovarian cancer often hinges on MDT collaboration, integrating expertise from gynecological oncologists, radiologists, pathologists, and ultrasound experts to interpret multimodal imaging data. However, access to MDTs is typically limited to tertiary or specialized centers. The patient diagnostic journey typically progresses from initial assessments by individual clinicians in primary care, determining referral to specialized institutions for further evaluation by gynecologists before MDT involvement. To assess the clinical utility of the OVUCM system, we compared its performance with that of both individual gynecologists and the Cancer Center MDT. Our preliminary results, based on a small sample of 35 patients, demonstrated that the OVUCM system achieved performance comparable to that of MDT consensus across most tasks (ΔAUC: 0.025–0.300; DeLong test, P > 0.05), while outperforming individual gynecologists in all tasks. This likely reflects the inherent complexity of ovarian cancer imaging and the high level of specialization required for accurate interpretation. These findings may indicate that, for multimodal image interpretation, even experienced gynecologists may encounter difficulties, a challenge likely to be even more evident among general practitioners. External validation of the OVUCM system in two non-specialized hospitals supported its potential generalizability, maintaining the AUC 0.833-0.974 without the need for subspecialty infrastructure. This capability has the potential to contribute to addressing a pressing global health challenge: fewer than 35% of ovarian cancer patients in non-tertiary centers receive guideline-concordant staging or surgery50. The OVUCM system was designed for embedding within medical information systems (e.g., PACS or radiology workstations), enabling automated retrieval and integration of multimodal imaging data. By delivering MDT-like diagnostic support in resource-limited settings, it could contribute to two critical clinical pathways. The first is early referral optimization. Accurate preoperative risk stratification facilitates timely referrals to gynecologic oncology centers, where specialized debulking surgery improves 5-year survival for patients by about 10-20%51–53. The second is diagnostic democratization, equipping frontline clinicians with standardized, interpretable risk assessments and reducing dependence on subjective expertise. The OVUCM system has several limitations that warrant consideration. First, as in many multimodal oncology studies, we faced the challenge of lacking large-scale and quality-controlled prospective datasets54. Our retrospective design, coupled with the relatively limited external validation, inherently introduces potential biases and the overfitting risk. Although we applied rigorous strategies (such as standardized data processing and validation) to enhance the robustness and generalizability, the prospective multicenter studies are warranted to further evaluate and optimize OVUCM’s performance. Second, precise segmentation of ultrasound images remains labor-intensive. Existing automated methods may be less reliable in tumors with irregular morphology or indistinct margins of malignant tumors. Our future work will focus on automated and human-in-the-loop segmentation techniques to improve workflow efficiency. In conclusion, we have developed the OVUCM system by integrating ultrasound, CT, and MRI features, which addresses the limitations of single-modality approaches and explores preoperative predictions across multiple disease outcomes. OVUCM was designed to unify multimodal imaging information into an interpretable, clinically aligned framework using data generated during standard clinical care, which may facilitate integration into existing workflows. This system demonstrates potential to complement MDT diagnosis-making, especially in settings with limited expertise, by supporting tumor treatment recommendations. This study highlights the feasibility of multimodal imaging-based artificial intelligence for personalized oncology and provides a foundation for future prospective and multicenter validation. Methods This study was approved by the Ethics Committee of Women’s Hospital, Zhejiang University School of Medicine (Approval number: IRB-20230405-R). Given the retrospective nature of this study and the International Ethical Guidelines of the Council for International Organizations of Medical Sciences55, a waiver of informed consent was obtained. Patients’ clinical records and imaging data were anonymized. Patients and data collection The internal datasets were retrospectively collected from patients with pathology-confirmed ovarian masses at the Women’s Hospital, Zhejiang University School of Medicine (WHZJU) between January 2014 and December 2023. They were established based on the following inclusion criteria: (1) surgery was the primary treatment, (2) recorded clinical variables of age, body mass index, gravidity, parity, menopause status, and FIGO Staging were available. (3) pelvic ultrasound, CECT, and MRI imaging within 4 weeks before surgery, (4) histological evidence from paraffin section of operative tissues. Additionally, external datasets were retrospectively collected from patients at ZJU4H between January 2019 and December 2024, and FHWMU between January 2022 and December 2024. Unlike the internal set, the external sets included patients based on the preoperative imaging requirements for the OVUCM system. Specifically, for Task n, the model required pelvic ultrasound, CECT, and MRI; therefore, patients who underwent all three imaging modalities were included. In a specific task where one or two modalities were sufficient, patients who completed the required images were considered. The following cases were excluded from both the internal and external datasets: (1) recurrence of borderline or malignant ovarian masses, (2) fine-needle aspiration puncture biopsy of ovarian masses, (3) preoperative chemotherapy, (4) indeterminate histological origin: uncertain correspondence between imaging findings and pathological diagnosis, (5) other synchronous malignancies, such as gastrointestinal cancers, breast cancer, cervical cancer, (6) undetectable adnexal masses in one or more imaging modalities, (7) imaging artifacts, due to metal artifacts or respiratory motion, (8) incomplete lesion borders, such as pelvic tilt or out-of-field masses. Clinical parameters were collected, including age, body mass index, gravidity, parity, menopausal status, FIGO Staging, and histological diagnosis. Serum tumor markers, including CA125, CA199, CA153, carcinoembryonic antigen, alpha-fetoprotein, and squamous cell carcinoma antigen, were recorded before a 1-month operation. If multiple tests of tumor markers were performed, the value closest to the surgery date was selected. For missing data, if more than 20% of tumor marker values were unavailable, the corresponding indicator was excluded56. If missing data accounted for less than 20%, the median value of that respective indicator was used for imputation57. Subsequently, data underwent preprocessing steps, including format standardization and normalization, to ensure consistency across variables. All patients were histologically diagnosed and classified according to the World Health Organization guidelines (2020)58. For patients diagnosed before 2020, any updates to the pathological classification required a review of the slides. Tumor staging was determined based on the FIGO criteria (2014)7, as staging had no changes in the 2018 version. Images of ultrasound, CECT, and MRI were retrieved from the Institutional Picture Archiving Communication System in the format of Digital Imaging and Communications in Medicine (DICOM). Ultrasound images included both grayscale and color Doppler modalities. Preoperative venous-phase CECT and MRI scans (T2WI and DWI sequences) were acquired. If multiple scans were available for a given modality, the one closest to the data of operation was selected. For a patient with bilateral or multiple lesions of identical histology, the lesion with the largest maximal diameter was analyzed. If there were different histological diagnoses, the lesion with more complex ultrasound features (based on IOTA ADNEX and O-RADS criteria) was selected18,47,59. The same mass was then matched across ultrasound, CT, and MRI to ensure consistency. Additionally, structured radiological features were extracted from CT reports, including the presence or absence of hepatic, splenic, or pancreatic metastases, lymph node involvement, peritoneal metastasis, and ascites. These were binary-coded (0 = absent, 1 = present). Image preprocessing Ultrasound, CECT, and MRI images were first converted from DICOM to Nifty format for subsequent processing. For color Doppler, three different color Spaces, RGB, YBR_FULL_422, and MONOCHROME2, were unified into the RGB color space to ensure consistent color representation. CECT and MRI images were resampled to voxel sizes of [1, 1, 1] mm³ and [1.5, 1.5, 7] mm³, respectively. To minimize intensity-related scale differences across imaging modalities and equipment, all images underwent Z-score normalization, standardizing voxel intensities to a mean of 0 and a standard deviation of 1. ROI segmentation Three ultrasound physicians with over 5 years of experience in obstetric and gynecologic ultrasound independently performed manual segmentation for the region of interest (ROI) on ultrasound images using ITK-SNAP software version 3.8.0 (open source, http://www.itksnap.org)60. The entire tumor boundary was traced on the grayscale images, while the solid components of the masses, including papillary projection and cystic septum, were delineated on color Doppler images61,62. To reduce inter-observer variability in manual annotations, final consensus masks were generated using the Simultaneous Truth and Performance Level Estimation (STAPLE) voting algorithm63. Lesion segmentation for CECT and MRI was performed using a deep learning-assisted semi-automatic pipeline, combining a modified volumetric nnU-Net64 automation with expert manual refinement. Two radiologists with over 5 years of experience in obstetric and gynecologic radiology manually delineated the lesion ROI on 30% of internal datasets using 3D Slicer software version 5.4.0 (open source, http://www.slicer.org)65. The segmentation model took CECT as a single-channel input, whereas both DWI and T2WI were used as a two-channel input. The trained model then segmented the remaining cases. This was followed by mandatory quality control: a senior ultrasound physician/radiologist (>15 years’ experience in obstetric and gynecologic imaging) reviewed and corrected ROI, especially in anatomically complex or discordant regions. All annotators and reviewers were blinded to the histological and clinical data. Radiomics feature extraction and selection We used the PyRadiomics66 toolkit to extract features from ultrasound, CECT, and MRI images, including shape, first order, and matrices such as gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), gray level dependence matrix (GLDM), and neighborhood gray tone difference matrix (NGTDM). For the ultrasound images, if multiple images per patient were available, we averaged their features. The missing grayscale or color Doppler scans were input by the median feature vector from cases sharing the same five clinical classification labels. Thus, each patient had 464-dimensional features for both color Doppler and grayscale images. For CECT and MRI (T2WI and DWI sequences) images, we extracted 1288-dimensional features from both original and filtered images with the Laplacian of Gaussian filter and Wavelet. A feature selection pipeline, which contained four steps including ANOVA67, Spearman correlation coefficient68, statistical test (Relief)69, and Lasso70, was applied to reduce redundancy by eliminating low-variance, highly correlated, or weakly predictive features, and shrinking irrelevant coefficients to zero. Single-modality model development Following feature filtering, four ranked feature sets were derived for each imaging modality (ultrasound, CECT, T2WI, DWI), with predefined feature set sizes at top k (k in [5, 10, 15, 20, 30, 40, 50]). Top k features were then input into a set of linear machine learning classifiers (Linear Regression, Logistic Regression, and Support Vector Classifier), and nonlinear classifiers (Naive Bayes and XGB), to develop the candidate models. Using a 5×4 nested cross-validation, the optimal model with the best feature number, classifier, and classifier hyperparameters was determined for each single-modality with respect to the highest AUC. The features corresponding to these optimal models were ranked by importance and formed the multi-modality feature pool for 11 modality combinations. Additionally, four radiological features extracted from CT reports were incorporated into the CT candidate feature pool. Ultrasound-CECT-MRI fusion The 11 possible combinations of modalities included US-CECT, US-DWI, US-T2WI, CECT-DWI, etc., up to US-CECT-DWI-T2WI. Following a series of multi-modality intermediate fusion steps (identical to the single-modality model training and validation process), the top k features (k in [5,10, 20, 30, 40, 50]) from the multi-modality feature pool, with five machine learning classifiers, form the candidate multi-modality models. The resulting candidate models from various combinations of dual-, triple-, and quadruple-modality configurations were trained and evaluated. The optimal model for each multi-modality combination was determined using 5 × 4 nested cross-validation based on the highest AUC values. From the aforementioned 4 optimal single-modality models and 11 optimal multi-modality models, the 15 optimal models for each task were determined. OVUCM system construction For each clinical task, the task-specific fusion model with the highest AUC among the 15 optimal models was selected for final integration. The significance of features was visualized through Permutation Importance Analysis to enhance model explainability and transparency of each model. The OVUCM system was then developed by incorporating clinical characteristics (e.g., age, menopausal status) and serum tumor markers (e.g., CA125) into the five selected task-specific models. This system enabled concurrent multi-task predictions while preserving modality-specific discriminative features. All models were developed using supervised learning algorithms, with ground truth labels available for each prediction task. Nested cross-validation We utilized a 5 × 4 nested cross-validation strategy71 (5-fold outer loop and 4-fold inner loop) for model training, validation, and internal testing. The internal dataset was partitioned into five subsets using stratified random sampling, preserving class distributions. For each task, four outer-loop subsets were used for training and inner-loop validation, utilizing hyperparameter optimization and model development, while the fifth subset served as an independent test. Within each outer loop iteration, the inner loop was used for hyperparameter tuning and model selection. The hyperparameters included the number of radiomics features and classifier configurations. We implemented the Synthetic Minority Oversampling Technique (SMOTE)72 to mitigate class imbalance, and GridSearchCV73 to select optimal hyperparameters. The mean performance metrics across five folds were reported as the test results. OVUCM system evaluation We performed an external evaluation using independent datasets from two general hospitals. Model calibration on these external cohorts was assessed using the Hosmer-Lemeshow test74, and recalibration was performed via Platt scaling75. Calibration curves76 were plotted to examine the concordance between predicted probabilities and actual outcomes, with the Brier score quantifying the overall prediction error. ROC curves were generated to evaluate the classification performance of the OVUCM system while visualizing the trade-off between sensitivity and specificity. Key evaluation metrics included AUC with its 95% CI, accuracy, precision, and F1 scores. The 95% CIs for AUCs were estimated using the bootstrapping77. Confusion matrices were generated to provide a detailed breakdown of classification results, including true positives, false positives, true negatives, and false negatives, for detailed performance interpretation. Decision curve analysis (DCA) was conducted to assess the net clinical benefit of the model across a range of threshold probabilities78. Comparison of the OVUCM system with clinicians’ performance We selected 35 patients from the internal datasets through a stratified random sampling. Random numbers were generated in Excel using the =RAND () function, and sorted to randomize patient order. The =RANK () function ensured unique rankings and prevented duplication. Sampling began with Task 5, from which 10 cases were drawn from the HGSOC, and five more cases were added per task to other datasets. This approach ensured a balanced distribution across all five binary classification tasks while preserving randomness and representativeness for comparison with clinical experts and the OVUCM system. We convened an MDT at Cancer Center, comprising three specialized experts: a chief gynecological oncologist, a chief ultrasound physician, and a chief radiologist. This MDT conducted a consensus review of the imaging data, engaging in structured deliberation on all five binary outcomes for the 35 patients. In parallel, three independent gynecologists with over 5 years of clinical experience individually assessed the same imaging datasets and recorded their assessments in standardized binary diagnostic tables across all five tasks. Diagnostic performance was quantified using standardized metrics: AUC and its 95% CI, accuracy, specificity, sensitivity, precision, and F1 scores. All clinicians were blinded to histological diagnoses and clinical outcomes. Comparative Study between OVUCM system, IOTA-ADNEX model, and O-RADS model Two ultrasound physicians (>5 and >10 years of experience) and two radiologists (>5 and >10 years) in obstetric and gynecologic imaging retrospectively evaluated an independent external ultrasound dataset from ZJU4H (n = 102) after standardized training in IOTA-ADNEX and O-RADS criteria. All evaluators were blinded to outcomes. Diagnostic performance was assessed using AUC, sensitivity, and specificity at all clinically relevant risk thresholds (3%, 5%, 10%, and 15% for IOTA-ADNEX and 1%, 10% and 50% for O-RADS). Statistical analysis Median (IQR) was used to evaluate continuous variables, while categorical variables were reported as absolute and percentages (%). Pairwise comparisons of proportions were conducted using two-sided McNemar tests, with statistical significance defined as P < 0.05. The Delong test compared the AUC of the OVUCM system and that of clinicians. The two-sided statistical significance threshold was set at α = 0.05. All computational workflows were implemented through an integrated software environment comprising SPSS Statistics (Version 26.0), Python (Version 3.7), and R (Version 4.3.0) programming languages. Supplementary information Supplementary information Supplementary information