A multicenter deep learning framework integrating radiomics and vision transformers for comprehensive ovarian tumor analysis from ultrasound imaging.

An overview of the proposed multi-task framework—covering tumor segmentation, feature extraction, classification, and PFS prediction—is shown in Fig.  1 . To ensure data quality and clinical relevance, several inclusion and exclusion criteria were applied: age-related exclusion: 1234 patients did not meet the specified age requirements. Incomplete or poor-quality imaging: 1145 patients lacked complete ultrasound scans or presented low-quality images. Non-tumor ovarian conditions: 854 patients were excluded due to pathologies unrelated to tumors (e.g., polycystic ovary syndrome, functional cysts). Unclassifiable tumors: 300 patients had tumors that could not be assigned to the six defined categories. The mean age of participants was 42.6 ± 10.3 years, representing a middle-aged female cohort consistent with the typical age range of ovarian tumor presentation. The dataset was dominated by benign tumors, including cystadenoma (39.4%), mature cystic teratoma (16.5%), and endometrioma (9.8%). Notably, malignant tumors accounted for 19.3%, providing sufficient representation for performance evaluation across diverse pathological conditions. Fig. 1 Overview of the study workflow and multitask framework Overview of the study workflow and multitask framework The mean tumor size was 5.4 ± 2.1 cm, reflecting variability in lesion volume and underscoring the need for segmentation models capable of handling heterogeneous morphologies. Tumor location was evenly distributed between the left and right ovaries, while 4.6% of cases involved both ovaries, highlighting anatomical diversity. Data were acquired using eight different ultrasound systems, with LOGIQ E9 and Philips IU22 contributing the largest share. This multi-platform variability supports the reproducibility and generalizability of model evaluation. Regarding clinical factors, PCOS (9.97%) and endometriosis (10.1%) were common comorbidities, both relevant for differential diagnosis. A positive family history of cancer was observed in 12% of patients, emphasizing the importance of early, accurate diagnosis in high-risk populations. The mean follow-up duration of 26.4 months provided robust longitudinal data, particularly valuable for PFS prediction. Collectively, these characteristics highlight the dataset’s diversity, clinical validity, and suitability for evaluating segmentation, classification, and prognostic modeling. Figures  2 , 3 and 4 present the segmentation performance of the five models on the training, internal test, and external test sets. Across all datasets and metrics, UNETR consistently outperformed competing architectures, demonstrating superior accuracy, robustness, and generalizability. Fig. 2 Segmentation performance on training set (mean ± SD) Fig. 3 Segmentation performance on internal test set Fig. 4 Segmentation performance on external test set Segmentation performance on training set (mean ± SD) Segmentation performance on internal test set Segmentation performance on external test set On the training set, UNETR achieved a Dice of 96.3% ± 1.2 and IoU of 93.1% ± 1.4, with low dHD (4.2%), ASSD (2.7%), and VOE (7.5%), confirming both geometric and volumetric precision. nnU-Net followed closely (Dice: 95.2%, IoU: 91.4%), validating its strong adaptability in multicenter data. As expected, performance slightly decreased on the internal and external test sets due to real-world variability, though the decline was marginal. On the external cohort, UNETR maintained high performance (Dice: 95.1%, IoU: 91.7%) with only small increases in ASSD and VOE, confirming its stability across heterogeneous imaging protocols. The relative ranking of models remained unchanged: Swin-UNet delivered competitive results (~ 93% Dice) but appeared more data-dependent, while SegNet and UNet, as classical CNN-based architectures, lagged in IoU and surface-based metrics, reflecting limitations in capturing complex ovarian tumor boundaries. Overall, UNETR emerged as the most effective model, consistently achieving high Dice/IoU with low boundary errors across all datasets. Its strong generalization to both internal and external cohorts underscores its clinical potential for robust ovarian tumor segmentation. Figure  5 illustrates the training dynamics of the five segmentation models in terms of DSC and loss over 250 epochs. The top row shows the evolution of DSC and loss during training, while the bottom row depicts the corresponding curves on the internal test set. UNETR and nnU-Net demonstrated faster convergence and greater stability than the other models, achieving higher DSC values and lower final loss. These learning curves highlight differences in training stability, overfitting tendencies, and overall efficiency across the architectures. Fig. 5 Training and test curves of DSC and loss for all segmentation models across 250 epochs Training and test curves of DSC and loss for all segmentation models across 250 epochs Figure  6 presents qualitative comparisons of segmentation outputs from the five evaluated models—UNETR, nnU-Net, Swin-UNet, SegNet, and UNet—using representative ultrasound samples. Ground truth annotations (solid blue) and predicted tumor masks (dashed red) are overlaid to enable direct assessment of anatomical fidelity, boundary alignment, and local discrepancies. Fig. 6 Visual comparison of segmentation masks generated by five deep learning models Visual comparison of segmentation masks generated by five deep learning models UNETR showed the closest agreement with reference contours, particularly in delineating irregular margins. nnU-Net produced consistent results overall but occasionally missed fine details in heterogeneous regions. Swin-UNet preserved structural continuity, though minor under-segmentation occurred near indistinct borders. In contrast, SegNet and UNet were more prone to misclassification at tumor peripheries, leading to over- or under-segmentation in low-contrast areas. Zoomed-in panels highlight challenging morphologies, visually confirming differences in segmentation quality. These qualitative findings emphasize the practical impact of segmentation accuracy on downstream classification and feature extraction. In the second stage of the pipeline, radiomic ( n  = 215) and deep features from ResNet ( n  = 2048) and ViT ( n  = 768) were extracted from tumor regions segmented by each model, yielding a combined 3031 features per segmentation model (Table  1 ). Table 1 Summary of feature reduction across models and methods Segmentation model Feature type Initial features After intraclass Correlation Coefficient (ICC) (> 0.75) After initial PCA RFE selected ANOVA selected UNETR Radiomic 215 178 – 65 52 ResNet (Deep) 2048 1610 – 110 98 ViT (Deep) 768 630 – 80 72 Combined features 3031 2470 250 130 115 nnU-Net Radiomic 215 172 – 63 50 ResNet (Deep) 2048 1585 – 108 95 ViT (Deep) 768 620 – 78 70 Combined features 3031 2437 245 128 112 Swin-UNet Radiomic 215 169 – 60 48 ResNet (Deep) 2048 1572 – 105 92 ViT (Deep) 768 615 – 76 68 Combined features 3031 2417 240 125 110 SegNet Radiomic 215 165 – 58 46 ResNet (Deep) 2048 1550 – 102 90 ViT (Deep) 768 610 – 75 66 Combined features 3031 2389 235 122 108 UNet Radiomic 215 162 – 55 44 ResNet (Deep) 2048 1535 – 100 88 ViT (Deep) 768 605 – 73 65 Combined features 3031 2362 230 120 106 Summary of feature reduction across models and methods To assess robustness across centers, ICC analysis was applied, retaining only features with ICC ≥ 0.75. Depending on segmentation quality, 2362–2470 reproducible features were preserved, with higher counts observed for UNETR and nnU-Net, consistent with their superior segmentation accuracy. Dimensionality was first reduced via PCA, compressing the feature space to ~ 230–250 components while preserving most variance. Two selection strategies were then applied: Recursive feature elimination (RFE): retained larger subsets (e.g., ~ 130 features for UNETR). ANOVA: selected fewer but statistically significant features. Across all methods, the UNETR-based pipeline consistently retained the largest number of informative features, reinforcing the link between high-quality segmentation and stable, discriminative feature extraction. Overall, this multi-stage process—from segmentation to reproducibility filtering, reduction, and selection—highlights how segmentation performance directly propagates into classification and prognostic prediction, ensuring downstream robustness. ICC analysis demonstrated high feature reproducibility across centers. Following conventional thresholds, ICC values ≥ 0.75 were considered indicative of good reproducibility, while ICC ≥ 0.90 denoted excellent reproducibility. Based on this criterion, the majority of retained radiomic descriptors and deep embeddings achieved at least good reproducibility, supporting their robustness for multicenter modeling. Quantitative evaluation of classification models revealed clear performance hierarchies across feature sets (Table  2 ). The combined feature set consistently achieved the highest metrics. Using TabTransformer with RFE, performance peaked with a training accuracy of 96.1%, sensitivity of 95.2%, and AUC of 98.0%, while maintaining strong generalizability on both the internal test set (accuracy: 95.2%, AUC: 96.9%) and the external cohort (accuracy: 94.0%, AUC: 95.8%). Table 2 Classification performance using UNETR-based features across feature types, classifiers, and datasets Feature type Feature selection Classifier Accuracy Sensitivity AUC Radiomic RFE XGBoost 90.6 ± 1.2/88.2/87.0 89.4 ± 1.4/87.0/85.4 93.1 ± 1.1/91.0/89.7 MLP 89.8 ± 1.5/87.4/86.2 88.5 ± 1.7/86.0/84.3 92.7 ± 1.4/90.3/88.9 TabTransformer 91.2 ± 1.3/88.7/87.6 90.1 ± 1.6/87.5/86.1 93.4 ± 1.2/91.5/90.5 ANOVA XGBoost 89.7 ± 1.4/87.0/85.9 88.3 ± 1.5/85.2/83.5 92.5 ± 1.3/89.9/88.0 MLP 88.9 ± 1.6/86.2/85.1 87.8 ± 1.7/84.9/83.0 91.6 ± 1.5/89.1/87.5 TabTransformer 90.3 ± 1.3/88.0/86.7 89.1 ± 1.5/86.3/85.2 92.8 ± 1.2/90.6/89.3 ResNet RFE XGBoost 94.5 ± 0.9/92.5/91.2 93.7 ± 1.1/91.2/89.6 96.8 ± 0.8/95.3/94.0 MLP 93.9 ± 1.0/91.8/90.6 93.0 ± 1.2/90.5/88.7 96.3 ± 0.9/94.5/93.1 TabTransformer 94.8 ± 0.8/93.0/91.9 94.1 ± 1.0/91.6/90.2 97.0 ± 0.7/95.7/94.6 ANOVA XGBoost 93.7 ± 1.0/91.5/90.4 92.8 ± 1.2/90.1/88.2 96.0 ± 0.9/94.2/92.6 MLP 93.1 ± 1.2/90.6/89.3 92.0 ± 1.3/89.2/87.1 95.4 ± 1.0/93.4/91.5 TabTransformer 94.0 ± 1.0/92.2/90.9 93.3 ± 1.1/90.9/89.4 96.5 ± 0.8/94.9/93.7 ViT RFE XGBoost 93.6 ± 1.0/91.3/90.0 92.8 ± 1.2/90.0/88.1 96.1 ± 0.9/94.0/92.6 MLP 92.9 ± 1.1/90.7/89.2 91.7 ± 1.3/89.2/87.3 95.5 ± 1.0/93.0/91.5 TabTransformer 93.8 ± 0.9/91.9/90.7 93.1 ± 1.1/90.7/89.0 96.3 ± 0.8/94.5/93.3 ANOVA XGBoost 92.7 ± 1.1/90.4/88.8 91.5 ± 1.3/89.3/87.2 95.3 ± 1.0/92.9/91.0 MLP 91.8 ± 1.2/89.5/87.9 90.4 ± 1.4/87.8/85.7 94.4 ± 1.1/92.1/90.0 TabTransformer 92.9 ± 1.0/90.9/89.5 91.9 ± 1.2/89.6/88.2 95.6 ± 0.9/93.7/92.2 Combined RFE XGBoost 96.4 ± 0.7/94.9/93.6 95.8 ± 0.8/93.7/92.4 98.1 ± 0.6/96.7/95.5 MLP 95.5 ± 0.9/94.1/92.8 94.7 ± 1.1/92.9/91.2 97.4 ± 0.8/96.0/94.6 TabTransformer 96.1 ± 0.8/95.2/94.0 95.2 ± 0.9/94.1/93.0 98.0 ± 0.7/96.9/95.76 ANOVA XGBoost 95.8 ± 0.8/94.5/93.3 94.9 ± 1.0/93.0/91.6 97.6 ± 0.7/96.4/95.1 MLP 95.0 ± 1.0/93.7/92.2 94.3 ± 1.1/92.4/90.7 97.2 ± 0.9/95.8/94.2 TabTransformer 96.2 ± 0.9/94.8/93.7 95.5 ± 1.0/93.7/92.5 97.9 ± 0.7/96.6/95.4 Each cell: training ± SD/internal test/external test (%) Classification performance using UNETR-based features across feature types, classifiers, and datasets Each cell: training ± SD/internal test/external test (%) Deep features extracted from ResNet ranked second (training AUC: 97.0%; external AUC: 94.6%), followed by ViT features (training AUC: 96.3%; external AUC: 93.3%). In contrast, radiomic features alone yielded the lowest performance (training AUC: 93.4%; external AUC: 90.5%), underscoring their limited discriminative capacity compared to deep or hybrid feature sets. Across all experiments, TabTransformer consistently outperformed MLP and XGBoost, particularly in combination with RFE, which emerged as the most effective selection method. Collectively, these results confirm the diagnostic advantage of integrating handcrafted radiomics with deep features and demonstrate the superior generalizability of transformer-based classifiers in multi-class ovarian tumor classification. The classification models derived from nnU-Net-based segmentation masks showed strong and consistent performance, though slightly lower than those achieved with UNETR features (Table  3 ). Among the feature types, the combined feature set again achieved the best overall results. Using RFE + TabTransformer, the combined set reached training accuracy of 94.5%, sensitivity of 92.3%, and AUC of 96.5%, while maintaining strong generalizability on both the internal test set (accuracy: 92.3%, AUC: 94.5%) and the external test set (accuracy: 90.8%, AUC: 93.0%). These findings highlight the benefit of integrating deep and handcrafted features for robust multi-class tumor prediction across heterogeneous datasets. Table 3 Classification performance using nnU-Net-based features across feature types, classifiers, and Datasets Feature type Feature selection Classifier Accuracy Sensitivity AUC Radiomic RFE XGBoost 88.2 ± 1.3/85.0/83.7 86.9 ± 1.5/84.0/82.1 90.5 ± 1.2/88.3/86.9 MLP 87.5 ± 1.4/84.2/82.9 85.8 ± 1.6/83.0/81.0 89.8 ± 1.3/87.5/86.0 TabTransformer 88.9 ± 1.2/85.5/84.3 87.4 ± 1.4/84.3/82.8 91.0 ± 1.1/89.0/87.5 ANOVA XGBoost 87.3 ± 1.4/83.8/82.5 85.6 ± 1.5/82.5/80.7 89.2 ± 1.3/86.9/85.2 MLP 86.7 ± 1.5/83.0/81.7 84.9 ± 1.6/81.5/79.8 88.5 ± 1.4/86.0/84.3 TabTransformer 88.0 ± 1.3/84.5/83.2 86.5 ± 1.5/83.0/81.4 90.0 ± 1.2/87.8/86.0 ResNet RFE XGBoost 92.3 ± 1.0/90.0/88.7 91.0 ± 1.2/88.5/87.0 94.5 ± 0.9/92.5/91.0 MLP 91.6 ± 1.1/89.2/87.9 90.2 ± 1.3/87.5/86.0 93.8 ± 1.0/91.5/90.0 TabTransformer 92.8 ± 0.9/90.5/89.3 90.5 ± 1.1/89.0/87.5 94.9 ± 0.8/93.0/91.8 ANOVA XGBoost 91.2 ± 1.1/88.8/87.5 89.8 ± 1.3/87.0/85.2 93.2 ± 1.0/90.8/89.0 MLP 90.5 ± 1.2/88.0/86.7 89.0 ± 1.4/86.0/84.0 92.5 ± 1.1/90.0/88.2 TabTransformer 91.7 ± 1.0/89.5/88.2 90.3 ± 1.2/88.0/86.5 93.7 ± 0.9/91.5/90.0 ViT RFE XGBoost 91.0 ± 1.1/88.5/87.2 89.5 ± 1.3/87.0/85.5 93.0 ± 1.0/90.8/89.0 MLP 90.3 ± 1.2/87.8/86.5 88.8 ± 1.4/86.0/84.2 92.3 ± 1.1/90.0/88.0 TabTransformer 91.5 ± 1.0/89.0/87.8 89.0 ± 1.2/87.5/86.0 93.5 ± 0.9/91.2/89.8 ANOVA XGBoost 89.8 ± 1.2/87.0/85.7 88.2 ± 1.4/85.5/83.5 91.5 ± 1.1/89.0/87.0 MLP 89.0 ± 1.3/86.2/84.8 87.5 ± 1.5/84.5/82.0 90.7 ± 1.2/88.0/86.0 TabTransformer 90.2 ± 1.1/87.5/86.2 88.7 ± 1.3/86.0/84.0 91.8 ± 1.0/89.5/87.5 Combined RFE XGBoost 94.0 ± 0.8/91.8/90.4 92.5 ± 1.0/90.0/88.5 96.0 ± 0.7/94.0/92.0 MLP 93.2 ± 1.0/90.7/89.1 91.6 ± 1.2/89.0/87.2 95.2 ± 0.9/93.0/91.1 TabTransformer 94.5 ± 0.8/92.3/90.8 92.3 ± 1.0/91.0/89.4 96.5 ± 0.6/94.48/93.04 ANOVA XGBoost 93.5 ± 0.9/91.0/89.6 91.9 ± 1.1/89.5/87.8 95.5 ± 0.8/93.5/91.5 MLP 92.6 ± 1.1/90.0/88.3 90.7 ± 1.3/88.0/86.1 94.7 ± 1.0/92.5/90.5 TabTransformer 93.8 ± 0.9/91.7/90.2 92.4 ± 1.1/90.0/88.3 96.0 ± 0.8/94.0/92.0 each cell: training ± SD/internal test/external test (%) Classification performance using nnU-Net-based features across feature types, classifiers, and Datasets each cell: training ± SD/internal test/external test (%) Deep features from ResNet ranked second (training AUC: 94.9%; external AUC: 91.8%), followed by ViT features (training AUC: 93.5%; external AUC: 89.8%). Radiomic features alone consistently showed the lowest performance (training AUC: 91.0%; external AUC: 87.5%), underscoring their limited discriminative power without complementary deep representations. Across all experiments, TabTransformer outperformed XGBoost and MLP, while RFE generally exceeded ANOVA in feature selection efficiency. Although a modest performance drop (3–5%) was observed from training to external testing, results remained within an acceptable range, confirming the robustness and generalizability of the nnU-Net feature-based classification pipeline. The classification results derived from Swin-UNet-based segmentation masks were slightly lower than those obtained with nnU-Net and UNETR features, yet still demonstrated consistent and generalizable performance across datasets (Table  4 ). The best results were obtained using combined features with RFE + TabTransformer, achieving a training accuracy of 93.0%, sensitivity of 90.9%, and AUC of 95.8%. On the internal test set, performance remained stable (accuracy: 90.9%, AUC: 93.7%), and external testing confirmed moderate robustness (accuracy: 89.4%, AUC: 92.1%), indicating that Swin-UNet supports reliable—though less competitive—classification outcomes compared to UNETR and nnU-Net. Table 4 Classification performance using Swin-UNet-based features across feature types, classifiers, and Datasets Feature type Feature selection Classifier Accuracy Sensitivity AUC Radiomic RFE XGBoost 86.5 ± 1.4/83.1/81.9 85.2 ± 1.6/82.0/80.2 88.7 ± 1.3/86.5/84.8 MLP 85.9 ± 1.5/82.3/81.0 84.5 ± 1.7/80.8/78.9 88.0 ± 1.4/85.0/83.5 TabTransformer 87.2 ± 1.3/84.0/82.7 86.0 ± 1.5/82.6/81.0 89.5 ± 1.2/87.2/85.6 ANOVA XGBoost 85.6 ± 1.5/81.5/80.1 84.0 ± 1.6/80.0/78.3 87.5 ± 1.4/85.0/83.0 MLP 85.0 ± 1.6/80.8/79.0 83.4 ± 1.7/78.9/77.1 86.9 ± 1.5/84.2/82.3 TabTransformer 86.4 ± 1.4/82.3/81.1 85.1 ± 1.5/81.0/79.5 88.2 ± 1.3/86.0/84.2 ResNet RFE XGBoost 90.8 ± 1.1/88.2/86.7 89.3 ± 1.3/86.5/85.1 93.0 ± 1.0/90.5/88.5 MLP 90.1 ± 1.2/87.4/85.9 88.6 ± 1.4/85.2/83.5 92.3 ± 1.1/89.6/87.0 TabTransformer 91.3 ± 1.0/88.9/87.3 88.9 ± 1.2/87.5/86.0 94.1 ± 0.9/91.5/89.3 ANOVA XGBoost 89.6 ± 1.2/86.5/85.0 88.0 ± 1.3/84.5/82.8 91.5 ± 1.0/89.0/87.0 MLP 88.9 ± 1.3/85.7/84.2 87.3 ± 1.4/83.8/81.9 90.7 ± 1.1/88.0/86.0 TabTransformer 90.2 ± 1.1/87.0/85.6 88.8 ± 1.2/85.2/83.7 92.5 ± 1.0/90.2/88.1 ViT RFE XGBoost 89.5 ± 1.1/86.7/85.3 88.0 ± 1.2/85.0/83.5 91.8 ± 1.0/89.2/87.3 MLP 88.7 ± 1.2/85.8/84.1 87.2 ± 1.4/83.7/81.8 91.0 ± 1.1/88.3/86.1 TabTransformer 90.0 ± 1.0/87.5/86.1 87.5 ± 1.2/86.0/84.4 92.7 ± 0.9/90.5/88.4 ANOVA XGBoost 88.3 ± 1.2/85.0/83.5 86.7 ± 1.3/83.2/81.5 90.5 ± 1.1/87.8/85.6 MLP 87.5 ± 1.3/84.2/82.4 86.0 ± 1.4/82.1/80.2 89.8 ± 1.2/86.6/84.5 TabTransformer 88.8 ± 1.1/85.8/84.3 87.5 ± 1.3/83.9/82.0 91.2 ± 1.0/88.7/86.6 Combined RFE XGBoost 92.7 ± 0.9/90.2/88.5 91.0 ± 1.1/88.4/87.0 95.0 ± 0.8/92.8/91.0 MLP 91.8 ± 1.0/89.3/87.7 90.3 ± 1.2/87.0/85.5 94.1 ± 0.9/91.5/90.0 TabTransformer 93.0 ± 0.8/90.9/89.4 90.9 ± 1.0/89.1/87.7 95.8 ± 0.7/93.68/92.1 ANOVA XGBoost 91.5 ± 1.0/88.5/87.0 90.0 ± 1.2/86.5/85.0 93.5 ± 0.9/91.0/89.2 MLP 90.8 ± 1.1/87.8/86.1 89.3 ± 1.3/85.2/83.6 92.7 ± 1.0/90.1/88.3 TabTransformer 92.2 ± 0.9/89.7/88.1 90.8 ± 1.1/87.8/86.3 94.6 ± 0.8/92.3/90.5 Each cell: training ± SD/internal test/external test (%) Classification performance using Swin-UNet-based features across feature types, classifiers, and Datasets Each cell: training ± SD/internal test/external test (%) Among individual feature groups, ResNet deep features ranked second, reaching training metrics of 91.3%/88.9%/94.1% and external test values of 87.3%/86.0%/89.3%. ViT features followed closely (training AUC: 92.7%; external AUC: 88.4%), though they exhibited slightly higher sensitivity to segmentation quality, as reflected in the larger performance drop (3–4%). In contrast, radiomic features alone consistently underperformed, with best training scores of 87.2%/86.0%/89.5% and external results declining to 82.7%/81.0%/85.6%, reaffirming their limited discriminative capacity without complementary deep features. Across all comparisons, TabTransformer paired with RFE again produced the strongest results, outperforming MLP, XGBoost, and ANOVA-based pipelines. Although Swin-UNet lagged behind UNETR and nnU-Net in absolute performance, the moderate degradation between training and external testing (4–6%) highlights its stability across diverse clinical data, supporting its use as a robust, albeit less optimal, segmentation backbone for classification tasks. The classification performance derived from SegNet-based segmentation masks ranked lowest among the evaluated backbones, though it maintained consistent feature-wise and classifier-wise trends (Table  5 ). The top-performing configuration was achieved using combined features with RFE + TabTransformer, yielding training metrics of 92.0%/90.9%/94.5%. On the internal test set, performance declined moderately (accuracy: 89.5%, sensitivity: 88.2%, AUC: 92.0%), while external validation confirmed a further decrease (accuracy: 87.8%, sensitivity: 86.5%, AUC: 89.9%), reflecting a generalization gap of ~ 4–5%—slightly higher than observed with stronger backbones. Table 5 Classification performance using SegNet-based features across feature types, classifiers, and datasets Feature type Feature selection Classifier Accuracy Sensitivity AUC Radiomic RFE XGBoost 84.7 ± 1.5/81.0/79.5 83.1 ± 1.6/79.5/77.8 87.0 ± 1.3/84.8/82.5 MLP 84.0 ± 1.6/80.2/78.3 82.4 ± 1.7/78.5/76.1 86.3 ± 1.4/84.0/81.2 TabTransformer 85.5 ± 1.4/82.0/80.6 83.9 ± 1.5/80.5/78.9 88.1 ± 1.2/85.7/83.8 ANOVA XGBoost 83.5 ± 1.6/79.0/77.4 81.8 ± 1.7/77.0/75.0 85.4 ± 1.4/83.0/80.9 MLP 82.9 ± 1.7/78.3/76.5 81.1 ± 1.8/76.3/74.1 84.6 ± 1.5/82.0/79.8 TabTransformer 84.3 ± 1.5/79.8/78.1 82.7 ± 1.6/78.0/76.2 86.0 ± 1.3/84.2/82.0 ResNet RFE XGBoost 89.0 ± 1.2/85.5/84.1 87.6 ± 1.3/84.0/82.3 91.2 ± 1.1/88.7/87.0 MLP 88.4 ± 1.3/84.6/83.2 86.8 ± 1.4/82.9/81.1 90.5 ± 1.2/87.9/86.0 TabTransformer 89.6 ± 1.1/86.2/84.9 88.2 ± 1.2/84.7/83.0 92.3 ± 1.0/89.5/88.0 ANOVA XGBoost 87.9 ± 1.3/83.5/82.0 86.2 ± 1.4/82.0/80.1 89.5 ± 1.2/87.0/85.5 MLP 87.2 ± 1.4/82.7/81.0 85.5 ± 1.5/80.9/78.9 88.8 ± 1.3/86.2/84.2 TabTransformer 88.5 ± 1.2/84.3/82.8 87.0 ± 1.3/83.0/81.2 90.6 ± 1.1/88.1/86.5 ViT RFE XGBoost 88.1 ± 1.2/84.6/83.0 86.6 ± 1.3/83.0/81.2 90.2 ± 1.1/87.6/85.7 MLP 87.4 ± 1.3/83.7/82.0 85.8 ± 1.4/81.9/79.8 89.4 ± 1.2/86.7/84.8 TabTransformer 88.8 ± 1.1/85.5/83.9 87.3 ± 1.2/84.0/82.3 91.0 ± 1.0/88.5/86.8 ANOVA XGBoost 86.5 ± 1.3/82.3/80.7 84.9 ± 1.4/81.0/78.9 88.4 ± 1.2/85.8/83.5 MLP 85.7 ± 1.4/81.4/79.8 84.2 ± 1.5/80.0/77.5 87.6 ± 1.3/84.6/82.3 TabTransformer 87.1 ± 1.2/83.0/81.5 85.5 ± 1.3/81.5/79.6 89.2 ± 1.1/86.7/84.5 Combined RFE XGBoost 91.5 ± 0.9/88.7/87.0 90.2 ± 1.1/87.0/85.5 93.6 ± 0.8/91.0/89.5 MLP 90.7 ± 1.0/87.8/86.0 89.0 ± 1.2/85.6/83.8 92.7 ± 0.9/89.8/88.2 TabTransformer 92.0 ± 0.8/89.5/87.8 90.9 ± 1.0/88.2/86.5 94.5 ± 0.7/92.03/89.92 ANOVA XGBoost 90.3 ± 1.0/86.8/85.2 88.6 ± 1.2/85.0/83.0 92.1 ± 0.9/89.6/87.5 MLP 89.5 ± 1.1/85.8/84.0 87.9 ± 1.3/83.8/82.0 91.3 ± 1.0/88.5/86.2 TabTransformer 90.8 ± 0.9/87.5/85.9 89.3 ± 1.1/86.0/84.0 93.0 ± 0.8/90.4/88.3 each cell: training ± SD/internal test/external test (%) Classification performance using SegNet-based features across feature types, classifiers, and datasets each cell: training ± SD/internal test/external test (%) Among individual feature sets, ResNet-derived features ranked second, with RFE + TabTransformer producing results of 89.6%/88.2%/92.3% in training and 84.9%/83.0%/88.0% in external testing. These values were clearly lower than the corresponding ResNet results under Swin-UNet (training AUC: 94.1%; external AUC: 89.3%), confirming a consistent hierarchical dependency of feature quality on segmentation backbone performance. ViT-based features also underperformed relative to more advanced backbones, peaking at 88.8%/87.3%/91.0% in training and 83.9%/82.3%/86.8% in external testing, compared to higher scores under Swin-UNet and nnU-Net. As with prior models, radiomic features alone yielded the weakest performance. Even under optimal conditions (RFE + TabTransformer), results were limited to 85.5%/83.9%/88.1% in training and 80.6%/78.9%/83.8% externally—lower than radiomic peaks achieved under Swin-UNet (AUC: 89.5%). Across all feature types, TabTransformer consistently outperformed MLP and XGBoost, and RFE again proved superior to ANOVA in selecting discriminative features. However, the slightly larger performance degradation from training to external testing (4–6%) highlights SegNet’s reduced stability and greater sensitivity to multicenter variability. In summary, SegNet, while functional, provides the weakest foundation for downstream classification, reinforcing the necessity of high-quality segmentation masks for maximizing diagnostic performance in ovarian tumor analysis. The classification performance derived from UNet-based segmentation masks was the lowest among the five evaluated backbones, reflecting UNet’s limited capacity to generate precise tumor boundaries (Table  6 ). The best overall results were obtained with combined features using RFE and TabTransformer, achieving training metrics of 90.8%/89.6%/93.3%. Performance declined moderately on the internal test set (accuracy: 88.0%, sensitivity: 86.5%, AUC: 91.0%) and further on the external test set (accuracy: 86.5%, sensitivity: 85.0%, AUC: 89.3%), indicating a generalization gap of ~ 4–5%—consistent but clearly lower than that of more advanced architectures. Table 6 Classification performance using UNet-based features across feature types, classifiers, and datasets Feature type Feature selection Classifier Accuracy Sensitivity AUC Radiomic RFE XGBoost 83.2 ± 1.5/79.5/78.1 81.7 ± 1.6/78.0/76.5 85.6 ± 1.3/83.5/81.4 MLP 82.5 ± 1.6/78.7/77.0 81.0 ± 1.7/76.8/74.9 84.9 ± 1.4/82.7/80.5 TabTransformer 83.8 ± 1.4/80.3/78.9 82.4 ± 1.5/79.0/77.2 86.5 ± 1.2/84.8/82.6 ANOVA XGBoost 81.8 ± 1.6/77.4/76.0 80.1 ± 1.7/75.5/73.8 83.9 ± 1.4/81.8/79.6 MLP 81.1 ± 1.7/76.5/74.9 79.5 ± 1.8/74.4/72.3 83.1 ± 1.5/80.9/78.4 TabTransformer 82.6 ± 1.5/78.2/76.7 81.2 ± 1.6/76.5/74.5 84.7 ± 1.3/82.5/80.2 ResNet RFE XGBoost 87.2 ± 1.2/83.8/82.5 85.9 ± 1.3/82.3/80.8 89.5 ± 1.1/86.9/85.3 MLP 86.5 ± 1.3/82.9/81.3 85.1 ± 1.4/81.0/79.1 88.8 ± 1.2/86.1/84.0 TabTransformer 87.7 ± 1.1/84.5/83.1 86.4 ± 1.2/83.0/81.2 90.2 ± 1.0/87.7/85.8 ANOVA XGBoost 85.8 ± 1.3/81.5/80.0 84.2 ± 1.4/79.8/77.5 87.6 ± 1.2/84.9/83.0 MLP 85.1 ± 1.4/80.7/78.9 83.6 ± 1.5/78.5/76.2 86.8 ± 1.3/84.0/82.1 TabTransformer 86.4 ± 1.2/82.1/80.7 85.0 ± 1.3/80.4/78.3 88.3 ± 1.1/85.5/83.5 ViT RFE XGBoost 86.3 ± 1.2/82.7/81.1 84.8 ± 1.3/81.0/79.4 88.4 ± 1.1/85.8/83.7 MLP 85.6 ± 1.3/81.8/80.0 83.9 ± 1.4/79.8/77.3 87.6 ± 1.2/84.9/82.5 TabTransformer 86.9 ± 1.1/83.5/81.9 85.5 ± 1.2/82.0/80.2 89.3 ± 1.0/86.7/84.9 ANOVA XGBoost 84.7 ± 1.3/80.2/78.8 83.2 ± 1.4/78.5/76.7 86.3 ± 1.2/83.5/81.4 MLP 84.0 ± 1.4/79.3/77.5 82.5 ± 1.5/77.0/75.0 85.6 ± 1.3/82.6/80.5 TabTransformer 85.3 ± 1.2/80.9/79.1 83.9 ± 1.3/79.5/77.4 87.1 ± 1.1/84.8/82.7 Combined RFE XGBoost 90.2 ± 0.9/87.4/85.9 88.9 ± 1.1/86.0/84.2 92.5 ± 0.8/89.8/88.0 MLP 89.5 ± 1.0/86.3/84.5 87.8 ± 1.2/84.5/82.6 91.6 ± 0.9/88.5/86.8 TabTransformer 90.8 ± 0.8/88.0/86.5 89.6 ± 1.0/86.5/85.0 93.3 ± 0.7/90.95/89.31 ANOVA XGBoost 88.9 ± 1.0/85.1/83.5 87.3 ± 1.2/83.5/81.5 91.0 ± 0.9/88.0/86.0 MLP 88.2 ± 1.1/84.0/82.2 86.5 ± 1.3/82.0/80.1 90.1 ± 1.0/87.0/85.0 TabTransformer 89.5 ± 0.9/85.8/84.1 88.0 ± 1.1/84.0/82.3 91.8 ± 0.8/88.9/87.0 Each cell: training ± SD/internal test/external test (%) Classification performance using UNet-based features across feature types, classifiers, and datasets Each cell: training ± SD/internal test/external test (%) Among deep feature groups, ResNet-derived features ranked second, with the optimal configuration (RFE + TabTransformer) achieving 87.7%/86.4%/90.2% in training, decreasing to 84.5%/83.0%/87.7% internally and 83.1%/81.2%/85.8% externally. This decline highlights the sensitivity of ResNet features to segmentation fidelity, confirming that UNet masks introduce greater noise into downstream pipelines compared to SegNet or Swin-UNet. ViT-based features followed, with peak values of 86.9%/85.5%/89.3% in training and 81.9%/80.2%/84.9% externally—approximately 2–3% lower than their Swin-UNet counterparts, again underscoring the segmentation–classification dependency. Radiomic features consistently showed the weakest discriminative power. Even under optimal settings (RFE + TabTransformer), performance was limited to 83.8%/82.4%/86.5% in training, dropping further to 78.9%/77.2%/82.6% in external validation. These results reaffirm that radiomics alone—especially when derived from less precise UNet masks—cannot provide reliable classification in complex multi-class scenarios. Across all feature categories, TabTransformer consistently outperformed MLP and XGBoost, while RFE surpassed ANOVA as the more effective feature selection strategy. However, the absolute performance across all UNet-derived pipelines remained inferior to those based on SegNet, Swin-UNet, nnU-Net, and UNETR. The 4–6% decline from training to external testing highlights the limited robustness of UNet features and reinforces the importance of higher-quality segmentation backbones for generalizable ovarian tumor classification in multicenter datasets. To streamline the presentation, we retained the ROC curves only for the best-performing segmentation backbone, UNETR, in the main text (Fig.  7 ). The corresponding ROC curves for the other segmentation models—nnU-Net, Swin-UNet, SegNet, and UNet—have been moved to the Supplementary Material (Figures S1–S4). Fig. 7 ROC curves for six-class classification using combined features from UNETR-based segmentation ROC curves for six-class classification using combined features from UNETR-based segmentation Figure  8 presents the confusion matrices illustrating the classification performance of the combined feature model with RFE and TabTransformer across the three datasets. The model demonstrates consistently high accuracy in distinguishing among tumor types, with overall accuracies of 96.1% (train), 95.2% (internal test), and 94.0% (external validation). Sensitivity values follow a similar trend (95.2%, 94.1%, and 93.0%, respectively), underscoring the robustness of the model in maintaining balanced detection performance across classes. Misclassifications are most frequently observed between histologically related entities such as cystadenoma and endometrioma, reflecting the inherent clinical and imaging overlap in these subtypes. Despite this, malignant tumors are correctly identified at a high rate, highlighting the clinical utility of the proposed approach in differentiating benign from malignant ovarian lesions across diverse cohorts. Fig. 8 Confusion matrices of the combined feature model with RFE and TabTransformer for train, internal test, and external validation datasets Confusion matrices of the combined feature model with RFE and TabTransformer for train, internal test, and external validation datasets Figure  9 provides a direct comparison of classification AUC across the five segmentation backbones. UNETR achieved the highest performance across all datasets (external test AUC: 95.76%), followed by nnU-Net (93.04%), Swin-UNet (92.10%), SegNet (89.92%), and UNet (89.31%). These results highlight the strong downstream impact of segmentation quality on classification, with transformer-based backbones (UNETR, nnU-Net, Swin-UNet) consistently outperforming conventional CNN-based models. Fig. 9 Comparison of AUC across segmentation backbones (UNETR, nnU-Net, Swin-UNet, SegNet, and UNet) using the combined feature set with RFE and TabTransformer. Bars represent AUC values for training, internal test, and external test datasets, with error bars showing mean ± SD for training Comparison of AUC across segmentation backbones (UNETR, nnU-Net, Swin-UNet, SegNet, and UNet) using the combined feature set with RFE and TabTransformer. Bars represent AUC values for training, internal test, and external test datasets, with error bars showing mean ± SD for training Figure  10 presents the two-dimensional distribution of the six ovarian tumor classes—cystadenoma, dermoid cyst, endometrioma, malignant tumors, normal ovarian tissue, and other benign tumors—projected into a transformed feature space. By merging radiomic signatures with high-level deep embeddings, refined through RFE, the visualization is classified using a TabTransformer. Fig. 10 Visualization of predicted class distributions for six ovarian tumor types across training, internal test, and external test sets Visualization of predicted class distributions for six ovarian tumor types across training, internal test, and external test sets In the plots, correctly classified samples are depicted as circles, whereas misclassified cases are indicated with crosses, providing an intuitive assessment of class separation and error patterns. The clustering of tumor subtypes highlights the ability of the integrated feature representation to capture both intra-class similarity and inter-class distinctiveness. The model achieved bootstrap-estimated accuracies of 96.1% (training), 95.2% (internal test), and 94.0% (external test), demonstrating robust generalization across independent datasets. Our results demonstrate that radiomic–deep feature integration enables accurate discrimination of complex ovarian tumor subtypes and strengthens the clinical utility of the proposed framework. To rigorously assess the observed differences in classification performance across segmentation backbones, DeLong’s test was applied to compare the AUC values obtained from the best-performing configuration (combined features + RFE + TabTransformer) for all five models (UNETR, nnU-Net, Swin-UNet, SegNet, and UNet). Pairwise comparisons demonstrated that the UNETR-based model consistently achieved significantly higher AUCs compared with all other backbones ( p  < 0.001 vs. UNet, p  = 0.002 vs. SegNet, p  = 0.011 vs. Swin-UNet, and p  = 0.038 vs. nnU-Net). These results confirm that the superior performance of UNETR is statistically robust and unlikely to be attributable to random variation. Similarly, nnU-Net significantly outperformed Swin-UNet, SegNet, and UNet ( p  < 0.05 in all comparisons). Differences between Swin-UNet and SegNet, as well as between SegNet and UNet, were also statistically significant but with smaller effect sizes, suggesting a gradual performance gradient that aligns with increasing architectural complexity. Overall, these findings confirm that the higher AUCs observed for UNETR and nnU-Net are not only numerically superior, but also statistically meaningful, underscoring the critical contribution of high-fidelity segmentation to the success of downstream multi-class ovarian tumor classification. The performance evaluation of the end-to-end classification models trained on cropped tumor regions derived from segmentation masks reveals consistent trends across segmentation backbones and datasets (Table  7 ). Overall, models trained on masks generated by UNETR and nnU-Net achieved the highest metrics, with ViT slightly outperforming ResNet in most cases. For instance, the ViT model trained on UNETR masks achieved the highest training accuracy (81.2%), sensitivity (80.6%), and AUC (81.9%), followed closely by nnU-Net. However, a modest performance drop is observed across the internal and external test sets, with external AUCs declining to 79.0% and 78.4% for UNETR and nnU-Net, respectively. Models based on Swin-UNet, SegNet, and especially UNet demonstrated lower classification performance, consistent with their comparatively less precise segmentation quality, which influenced the quality of cropped image regions used in the end-to-end training. Table 7 Performance of end-to-end classification models (ViT and ResNet) using cropped images from segmentation masks Segmentation model Classification model Dataset Accuracy (%) Sensitivity (%) AUC (%) UNETR ViT Training 81.2 ± 0.9 80.6 ± 1.1 81.9 ± 0.8 Internal test 79.4 78.5 80.2 External test 78.1 77.0 79.0 ResNet Training 80.6 ± 1.0 79.8 ± 1.2 81.0 ± 0.9 Internal test 78.7 77.5 79.5 External test 77.3 76.0 78.2 nnU-Net ViT Training 80.8 ± 0.8 80.0 ± 1.1 81.5 ± 0.7 Internal test 78.9 77.8 79.7 External test 77.5 76.3 78.4 ResNet Training 80.2 ± 0.9 79.3 ± 1.1 80.8 ± 0.8 Internal test 78.2 77.0 78.9 External test 76.8 75.6 77.5 Swin-UNet ViT Training 80.0 ± 0.8 79.2 ± 1.0 80.6 ± 0.7 Internal test 78.1 77.0 78.8 External test 76.7 75.4 77.6 ResNet Training 79.5 ± 1.0 78.5 ± 1.2 80.0 ± 0.9 Internal test 77.6 76.3 78.2 External test 76.0 74.8 76.9 SegNet ViT Training 79.4 ± 0.9 78.6 ± 1.0 80.1 ± 0.8 Internal test 77.5 76.3 78.3 External test 76.1 74.9 77.0 ResNet Training 78.8 ± 1.1 77.8 ± 1.3 79.4 ± 1.0 Internal test 76.9 75.5 77.6 External test 75.4 73.9 76.2 UNet ViT Training 78.7 ± 1.0 77.5 ± 1.1 79.5 ± 0.9 Internal test 76.8 75.6 77.4 External test 75.3 74.1 76.1 ResNet Training 78.0 ± 1.1 76.9 ± 1.3 78.8 ± 1.0 Internal test 76.1 74.8 76.9 External test 74.6 73.1 75.5 Performance of end-to-end classification models (ViT and ResNet) using cropped images from segmentation masks Additionally, ViT models consistently outperformed ResNet across all segmentation inputs, particularly in sensitivity and AUC, underscoring the benefits of transformer-based architectures in handling contextual information even when trained directly on localized tumor crops. End-to-end models, while competitive, consistently underperformed relative to the feature-based pipeline that fused radiomics with deep features, refined through feature selection and conventional classifiers. This suggests that structured feature engineering combined with dimensionality reduction provides greater discriminative power for multi-class ovarian tumor classification than direct end-to-end image-based learning in this context. Figures  11 and 12 present the classification training dynamics for five segmentation backbones—UNETR, nnU-Net, Swin-UNet, SegNet, and UNet—when paired with either a ViT or ResNet classifier in an end-to-end pipeline. Fig. 11 Training accuracy curves over 250 epochs for five segmentation models using ViT and ResNet classifiers Fig. 12 Training loss curves over 250 epochs for five segmentation models using ViT and ResNet classifiers Training accuracy curves over 250 epochs for five segmentation models using ViT and ResNet classifiers Training loss curves over 250 epochs for five segmentation models using ViT and ResNet classifiers Figure  11 shows the accuracy curves over 250 epochs for both the training and internal test sets. ViT-based models generally exhibited faster initial accuracy gains, but with greater variability across backbones, whereas ResNet-based models showed more gradual but stable improvements. Figure  12 illustrates the corresponding loss trajectories, where ViT-based configurations consistently reached lower final loss values, indicating more effective optimization and stronger feature discrimination compared to ResNet. Together, these curves highlight the trade-off between rapid learning dynamics (ViT) and training stability (ResNet), providing insights into how classifier choice interacts with segmentation backbones in shaping downstream performance. The regression outcomes for PFS prediction using the optimal configuration—combined features extracted from UNETR-based segmentation and refined via RFE—are summarized in Table  8 . Table 8 Performance of regression models for PFS Prediction using UNETR-based combined features (RFE) Model Dataset MAE (months) RMSE (months) R 2 score C -index XGBoost Training 3.42 ± 0.21 4.65 ± 0.32 0.781 ± 0.018 0.810 Internal test 3.68 4.81 0.762 0.797 External test 3.89 4.96 0.747 0.784 MLP Training 3.56 ± 0.25 4.79 ± 0.34 0.763 ± 0.021 0.799 Internal test 3.83 4.92 0.749 0.786 External test 4.01 5.08 0.732 0.773 TabTransformer Training 2.95 ± 0.16 4.12 ± 0.22 0.833 ± 0.012 0.847 Internal test 3.18 4.28 0.807 0.828 External test 3.36 4.47 0.791 0.816 Bold values indicate the best-performing model for each metric and dataset Performance of regression models for PFS Prediction using UNETR-based combined features (RFE) Bold values indicate the best-performing model for each metric and dataset Across all datasets, the TabTransformer consistently outperformed XGBoost and MLP on every evaluation metric: Prediction accuracy: TabTransformer achieved the lowest MAE (2.95 ± 0.16 months in training, 3.18 in internal test, and 3.36 in external test), demonstrating minimal deviation from ground truth values. Error robustness: it also yielded the lowest RMSE (4.12 ± 0.22, 4.28, and 4.47 months), highlighting its ability to minimize large prediction errors. Model fit: TabTransformer achieved the highest R 2 values (0.833 ± 0.012 training, 0.807 internal, 0.791 external), explaining more than 79% of PFS variance even in external validation—evidence of strong generalization. Survival ranking: its C -index values (0.847, 0.828, and 0.816) confirmed robust discriminative capacity, crucial in survival analysis for maintaining correct patient risk stratification. By contrast, XGBoost and MLP achieved lower accuracy and generalization, with external R 2 values dropping to 0.747 and 0.732, and C -indices of 0.784 and 0.773, respectively. While both showed clinical utility, their higher error margins and reduced concordance suggest less stable predictive capacity. In summary, TabTransformer provided the most accurate, stable, and clinically interpretable PFS predictions, making it a strong candidate for integrated diagnostic–prognostic pipelines in ovarian tumor analysis. To further assess the prognostic utility of the regression model outputs, we conducted Kaplan–Meier survival analysis based on predicted PFS values from the best-performing model (TabTransformer with combined + RFE features). Patients were stratified into high-PFS and low-PFS groups using the median predicted PFS value of 26 months as the cutoff. This median-based approach ensured balanced group sizes and meaningful survival comparisons. The resulting Kaplan–Meier curves revealed clear and consistent separation between the two groups, demonstrating that the model effectively discriminated between patients with favorable versus unfavorable progression outcomes. This indicates that the TabTransformer is capable of delivering clinically actionable prognostic stratification in addition to diagnostic classification, supporting its role in a unified imaging-driven framework. To maintain conciseness, the Kaplan–Meier survival analyses have been moved to the Supplementary Material (S5).

This multicenter study included ultrasound imaging data from 8736 patients evaluated for ovarian tumors across eight medical centers. Incomplete follow-up or insufficient clinical data: 1133 patients were removed due to missing tumor information or lack of follow-up. After applying these criteria, the final dataset comprised 3156 patients. Tumors were categorized into six groups: cystadenoma, mature cystic teratoma (dermoid cyst), endometrioma, malignant tumors, normal ovarian tissue, and other benign tumors (e.g., fibromas, thecomas). Eligible participants were female patients aged 18 years or older who underwent ultrasound examination for ovarian tumors at one of the eight participating centers. Only cases belonging to six predefined categories—cystadenoma, mature cystic teratoma (dermoid cyst), endometrioma, malignant tumors, normal ovarian tissue, or other benign tumors—were included. Patients were required to have complete, good-quality ultrasound images, with tumor diagnoses confirmed by histopathology or clinical evaluation. Patients were excluded if they: (i) were younger than 18 years; (ii) had missing, incomplete, or poor-quality imaging data; (iii) lacked follow-up information; (iv) had non-tumor ovarian conditions (e.g., polycystic ovary syndrome, functional cysts); (v) presented with tumors not matching the six predefined categories; or (vi) had incomplete clinical or pathology records preventing reliable confirmation of tumor type or malignancy status. Table S1 presents detailed patient characteristics, and Table S2 summarizes the specifications of the ultrasound imaging systems used across centers; both are provided in the Supplementary Material. Ultrasound data were collected from eight medical centers, each employing different imaging systems. The specifications of the ultrasound devices used at each center are summarized in Table S2. This multicenter and device-diverse dataset provides a valuable foundation for robust evaluation of the proposed framework and enhances the generalizability of the results across heterogeneous clinical environments. Preprocessing was performed to enhance ultrasound image quality, reduce variability across imaging systems, and facilitate accurate tumor segmentation and classification. The workflow consisted of two main steps: image preprocessing and data augmentation. The original ultrasound images underwent several preprocessing operations, including noise reduction, contrast enhancement, and intensity normalization. Gaussian filtering was applied to suppress speckle noise and smooth irregularities. Histogram equalization was used to enhance contrast and improve tumor boundary visibility. Finally, pixel intensity normalization standardized image ranges across different ultrasound devices and acquisition settings, ensuring consistency within the dataset. To enhance robustness and reduce overfitting, data augmentation was applied during training. Augmentations included random rotations (± 15°), horizontal and vertical flips, scaling (0.9–1.1), and slight intensity shifts. All transformations preserved tumor morphology to ensure biological plausibility. To achieve accurate delineation of ovarian tumors in ultrasound images, five state-of-the-art deep learning architectures were implemented. These models were chosen for their demonstrated success in medical image segmentation and their ability to handle variability in tumor morphology and imaging conditions. UNet, a convolutional encoder–decoder network, served as the baseline model. Its architecture enables simultaneous capture of global structure and fine details, making it highly effective for biomedical segmentation tasks. The self-configuring nnU-Net automatically adapts its architecture, preprocessing, and training protocols to the characteristics of the input data. This adaptability ensures robust and generalizable performance across heterogeneous multicenter datasets without manual tuning. UNETR replaces conventional convolutions with transformer modules to model long-range spatial dependencies. By leveraging global contextual relationships, it excels at segmenting tumors with irregular or complex boundaries. Swin-UNet integrates Swin Transformers into a UNet-like architecture, capturing both local detail and hierarchical global context. Its multiscale representation improves accuracy in detecting subtle patterns and complex morphologies. SegNet employs an encoder–decoder framework that reuses max-pooling indices during decoding to preserve spatial detail. This design enhances performance in noisy ultrasound data and enables precise boundary delineation. All models were rigorously trained and evaluated on the multicenter dataset. Together, their complementary strengths established a comprehensive segmentation framework, effectively addressing the heterogeneity and complexity of ultrasound imaging in ovarian tumor analysis. All segmentation networks were trained under a standardized protocol to ensure fair comparison and consistent evaluation. Each model was trained for 250 epochs with an early stopping criterion to prevent overfitting. A batch size of 8 was used, balancing computational efficiency with GPU memory constraints. The learning rate was fixed at 1 × 10 − 4 , and optimization was performed using the Adam optimizer ( β ₁ = 0.9, β ₂ = 0.999). Data augmentation strategies, including random rotations, flips, scaling, and cropping, were applied during training to increase variability and improve generalization across heterogeneous multicenter imaging data. Model optimization was guided by the Dice loss, defined as: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{Dice Loss}} = 1 - \frac{{2\left| {A \cap B} \right|}}{\left| A \right| + \left| B \right|},$$\end{document} Dice Loss = 1 - 2 A ∩ B A + B , where A and B denote the predicted and ground truth tumor masks, respectively. Dice loss is particularly effective for highly imbalanced medical images, where the tumor region constitutes only a small fraction of the overall image. Segmentation performance was assessed using multiple complementary metrics: Dice similarity coefficient (DSC): \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{DSC}} = \frac{{2\left| {A \cap B} \right|}}{\left| A \right| + \left| B \right|}.$$\end{document} DSC = 2 A ∩ B A + B . Measures overlap between predicted and ground truth regions (0 = no overlap, 1 = perfect). Intersection over Union (IoU): \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{IoU}} = \frac{{\left| {A \cap B} \right|}}{{\left| {A \cup B} \right|}}.$$\end{document} IoU = A ∩ B A ∪ B . Quantifies the proportion of overlap relative to the total combined area. Values closer to 1 indicate better segmentation accuracy. Directed Hausdorff distance (dHD): \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{dHD}}\left( {A.B} \right) = \mathop {\max }\limits_{a \in A} \mathop {\min }\limits_{b \in B} \left| {\left| {a - b} \right|} \right|.$$\end{document} dHD A . B = max a ∈ A min b ∈ B a - b . Captures the maximum boundary discrepancy between predicted and reference masks. Average symmetric surface distance (ASSD): \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{ASSD}} = \frac{1}{N}\mathop \sum \limits_{i = 1}^{N} \left( {{\mathrm{dist}}\left( {A_{i} .B} \right) + {\mathrm{dist}}\left( {B_{i} .A} \right)} \right).$$\end{document} ASSD = 1 N ∑ i = 1 N dist A i . B + dist B i . A . Provides an average measure of boundary proximity. Volume overlap error (VOE): \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{VOE}} = \frac{{\left| {A \cup B} \right| - \left| {A \cap B} \right|}}{{\left| {A \cup B} \right|}}$$\end{document} VOE = A ∪ B - A ∩ B A ∪ B represents volumetric disagreement, where lower values indicate better overlap. From the segmented regions of interest (ROIs), we extracted both radiomic descriptors and deep learning features across the six tumor categories. These features served as inputs for subsequent dimensionality reduction and multi-class classification models. Radiomic features were computed from each ROI using a validated platform compliant with the Image Biomarker Standardization Initiative (IBSI) guidelines (SERA). In total, 215 radiomic features were extracted: 79 first-order descriptors (capturing tumor shape, size, and intensity distributions) and 136 higher-order 3D texture features derived from gray-level co-occurrence, run-length, and size-zone matrices. While first-order descriptors quantified basic tumor structure, higher-order features characterized intra-tumoral heterogeneity and complex spatial patterns. These features were evaluated both individually and in combination, and subsequently integrated into dimensionality reduction and classification pipelines. Deep feature extraction leveraged pre-trained ResNet-50 and ViT models applied to the segmented tumor regions. ResNet-50. Features were extracted from the penultimate convolutional and final fully connected layers, yielding 2048 features per sample, representing both low- and high-level structural characteristics. ViT. Features were derived from the output of the final transformer block, resulting in 768 features per sample, which captured long-range dependencies and global contextual relationships within tumor regions. The ViT model was configured with a patch size of 16 × 16 pixels, consistent with the original architecture for medical image analysis. Each segmented tumor region was resized and embedded into fixed-length token sequences corresponding to this patch size, enabling the transformer to effectively capture spatial dependencies at a moderate granularity. Radiomic and deep features were used independently and in combination, providing a comprehensive tumor characterization by uniting detailed texture-based descriptors with high-level spatial and contextual information. These feature sets were subsequently subjected to dimensionality reduction and classification. To ensure the reliability of extracted features across heterogeneous imaging systems and clinical settings, ICC analysis was performed prior to feature selection and dimensionality reduction. We evaluated feature reproducibility by computing ICC values separately for handcrafted radiomic biomarkers and transformer/CNN-derived deep features, based on ROIs from different centers. Features with ICC ≥ 0.75 were considered highly reproducible and retained for further analysis, while those with lower values were discarded due to their variability and potential to introduce noise. This filtering step ensured that only stable, cross-center features were included in downstream pipelines, thereby improving the robustness, generalizability, and clinical applicability of the classification and prognostic models. After confirming feature reproducibility with ICC analysis, multiple dimensionality reduction and feature selection strategies were applied to optimize the feature space and improve model efficiency. Specifically, principal component analysis (PCA), recursive feature elimination (RFE), and analysis of variance (ANOVA) were employed. For combined radiomic–deep feature sets, PCA was first used to compress the high-dimensional input while retaining most of its variance. The PCA-reduced features were then refined using RFE, which iteratively removed non-informative features based on model performance, and ANOVA, which identified statistically significant features for distinguishing tumor classes. The optimized feature sets were subsequently classified using three machine learning models: XGBoost, for its effectiveness with high-dimensional data and iterative boosting of weak learners; Multi-layer perceptron (MLP), for capturing complex non-linear relationships; The TabTransformer is a deep learning architecture designed for tabular data, which applies self-attention mechanisms to learn contextual relationships between features. Unlike conventional multilayer perceptrons, it captures inter-feature dependencies and hierarchical representations, making it particularly effective for structured datasets such as radiomic and deep feature combinations. These classifiers were trained to categorize tumors into six classes: cystadenoma, mature cystic teratoma, endometrioma, malignant tumors, normal ovarian tissue, and other benign tumors. The integration of robust feature reduction methods with diverse classification algorithms ensured accurate and generalizable ovarian tumor classification across the multicenter dataset. To extend the framework beyond tumor classification, PFS prediction was incorporated as a regression task. Features derived from radiomic analysis and from deep learning embeddings of the segmented tumor masks underwent the same preprocessing steps applied in classification, including ICC filtering (ICC ≥ 0.75), dimensionality reduction, and feature selection using RFE and ANOVA, thereby enhancing both reproducibility and computational efficiency. The reduced feature sets were then used to train three regression models—XGBoost, MLP, and TabTransformer—adapted to output continuous PFS estimates rather than discrete class labels. Model performance was assessed using standard regression and survival analysis metrics: Mean absolute error (MAE): average magnitude of prediction error. Root mean squared error (RMSE): penalizes large errors, reflecting overall prediction accuracy. R 2 (coefficient of determination): proportion of variance in PFS explained by the model. Concordance index ( C -index): measures the model’s ability to correctly rank patients by risk. For prognostic validation, Kaplan–Meier survival curves were generated by dividing patients into high- and low-risk groups based on the median predicted PFS. Significant separation between curves confirmed the prognostic value of the regression models. By integrating PFS prediction into the same pipeline as segmentation and classification, the proposed framework provides both diagnostic and prognostic insights, offering a comprehensive solution for ovarian tumor analysis in clinical practice. For all tasks—segmentation, classification, and PFS prediction—the dataset was split into 70% training and 30% internal testing subsets. An external validation cohort of 756 patients, excluded from training and internal testing, was used to provide an unbiased assessment of real-world performance. A comprehensive grid search was applied to maximize performance: For segmentation models, we explored learning rates ranging from 0.00001 to 0.001, with a fixed batch size of 16 and 250 training epochs. Optimal configurations were selected based on validation accuracy and training stability. XGBoost: number of trees (100–1000). MLP: 2–5 hidden layers with varying units. TabTransformer: number of attention heads and related parameters. All classifiers were trained with learning rates in the range of 0.00001 and 0.001. In addition to feature-based approaches, end-to-end ViT and ResNet models were implemented for tumor classification. Segmentation masks were used to crop tumor regions directly from the original ultrasound images, ensuring that only relevant areas were processed. The cropped regions were then fed directly into the networks, bypassing manual feature extraction. Both models were trained for 250 epochs and fine-tuned to classify tumors into six categories. During training, ResNet and ViT automatically learned spatial and texture-based representations from the cropped inputs, enabling end-to-end optimization and reducing reliance on traditional feature engineering. To ensure comparability between modeling strategies, the end-to-end ViT and ResNet classifiers were trained using the same training and validation splits as the feature-based pipelines. These models received tumor-cropped regions as input and were evaluated using the same external test set and standard performance metrics (accuracy, sensitivity, AUC), enabling consistent benchmarking across approaches. The results of this comparative evaluation are detailed in the Results section. Model performance was evaluated using accuracy, sensitivity (Recall), and the area under the ROC curve (AUC). Accuracy quantified the overall proportion of correctly classified cases. Sensitivity measured the ability to correctly detect true positives, which is particularly critical for early malignant tumor detection. AUC assessed the models’ discriminative ability across varying thresholds. To compare models, DeLong’s test was applied for pairwise AUC comparisons. In addition, 95% confidence intervals (CIs) were computed for accuracy, sensitivity, and AUC to evaluate the reliability and variability of performance estimates. All models were implemented in Python, utilizing TensorFlow and PyTorch for deep learning. NumPy and Pandas were used for data preprocessing, while scikit-learn supported feature selection, dimensionality reduction, and machine learning classifiers. SERA was employed for radiomic feature extraction in compliance with IBSI standards. Visualization and performance evaluation were performed using Matplotlib and Seaborn. Model training and evaluation were conducted on high-performance computing hardware, including NVIDIA Tesla V100 and RTX 3090 GPUs, supported by 16-core CPUs and 64 GB RAM. A distributed computing environment enabled efficient training of large-scale models on multicenter datasets.

In this study, we developed and validated a unified deep learning–radiomics framework for multicenter ultrasound-based ovarian tumor analysis. By combining handcrafted radiomic descriptors with deep learning representations, the approach yielded high-performance metrics across segmentation, multi-class classification, and progression-free survival prediction. The framework delivered both diagnostic and prognostic insights, highlighting its ability to reduce observer dependence and enhance reproducibility across institutions. Overall, the proposed framework shows strong potential for real-world implementation. It offers a scalable tool to support early detection, risk stratification, and personalized therapeutic planning in ovarian cancer management.

This study introduced and validated an innovative multi-task framework that integrates deep learning and radiomics for comprehensive ovarian tumor analysis, encompassing segmentation, classification, and PFS prediction. By leveraging a large-scale multicenter ultrasound dataset, our framework integrates handcrafted radiomic features with deep representations extracted from segmentation-guided models. Reproducibility of extracted features was systematically evaluated using the ICC, and robust feature selection techniques were applied to optimize model performance. Compared to similar studies in ovarian tumor classification, our framework achieved higher external validation performance, particularly in terms of AUC and sensitivity. For example, prior radiomics-only approaches often reported limited generalizability, with AUC values dropping substantially when applied to external datasets, likely due to feature instability across ultrasound vendors and acquisition protocols. Similarly, earlier deep learning studies without multicenter training frequently showed strong internal performance but weaker external reproducibility, reflecting overfitting to single-institution characteristics. By contrast, our integration of reproducible radiomic descriptors with Vision Transformer–based embeddings, combined with rigorous feature harmonization and multicenter data, helped maintain robust performance across internal and external cohorts. These differences underscore the importance of both dataset diversity and hybrid feature integration in achieving clinically reliable generalization. Among the segmentation architectures, UNETR demonstrated the highest performance in both segmentation accuracy and subsequent classification and prognostic tasks, followed by nnU-Net, Swin-UNet, SegNet, and UNet. Feature fusion with RFE and TabTransformer delivered the strongest results, yielding AUCs of 98.0% (training) and 95.76% (external test). For PFS prediction, the unified feature set achieved a C-index of 0.847, underscoring the prognostic reliability of the model. Collectively, these results highlight the clinical utility and generalizability of the proposed multi-task framework for ovarian tumor analysis across diverse imaging environments. In comparison to Du et al. [ 39 ], who proposed a deep learning radiomics nomogram (DLR_Nomogram) for binary classification of ovarian tumors, our framework extends the task to multi-class classification with six distinct tumor categories. While Du et al. [ 39 ] reported AUCs of 0.985 (training) and 0.928 (testing), our model achieved similarly high classification accuracy in a more complex diagnostic setting, demonstrating improved diagnostic precision without compromising performance. In contrast to their reliance on logistic regression for final prediction, our model leverages deep learning classifiers, which demonstrated superior adaptability to heterogeneous features. Zuo et al. [ 40 ] developed statistical radiomics-based prognostic models for ovarian cancer and reported C-indices between 0.773 and 0.794 for overall survival (OS) and recurrence-free survival (RFS). Our approach improves upon this by utilizing a unified feature pipeline derived from segmentation outputs and achieving higher C-indices (up to 0.847) for PFS. Importantly, our study includes validation on an independent external dataset, underscoring the generalizability of our prognostic modeling. In another study by Du et al. [ 41 ], a three-class model was proposed to classify benign, borderline, and malignant tumors, with micro- and macro-average AUCs of 0.90 and 0.84, respectively. Although comparable, our model expands this taxonomy to six classes and achieves per-class ROC values exceeding 0.90 in most cases, indicating improved diagnostic precision. Moreover, we addressed the common limitation in their study—difficulty in recognizing borderline tumors—through a combined radiomic–deep learning strategy. Jin et al. [ 42 ] focused on the reproducibility of radiomic features derived from various U-Net architectures. While they confirmed the validity of automatic segmentation for feature stability, our study expanded this analysis by applying ICC filtering to features extracted from five segmentation models and demonstrating how segmentation quality influences downstream classification and regression tasks. This not only confirms reproducibility, but also establishes a direct link between segmentation fidelity and clinical outcomes. Wu et al. [ 37 ] demonstrated the power of CNNs in histologic subtype classification using ultrasound images, achieving an accuracy of 95.2% with ResNext50. While their approach emphasized histological differentiation, our study focused on real-world diagnostic classes and incorporated both radiomic and deep feature sets, enhancing interpretability and feature relevance in clinical practice. Hsu et al. [ 6 ] developed an ensemble CNN system for ovarian tumor classification with mean accuracies of 92.15% and sensitivity/specificity above 91%. However, their framework lacked segmentation-guided preprocessing and feature reproducibility analysis. Our pipeline integrates segmentation, feature engineering, and interpretability mechanisms, including ICC-based filtering and Kaplan–Meier analysis, which are essential for clinical translation. Finally, Arezzo et al. [ 43 ] utilized ultrasound-based machine learning to predict 12-month PFS in ovarian cancer, reporting AUROC of 0.92 using Random Forest. Our regression approach not only achieved similar or better predictive performance but extended the horizon of prediction to longer-term PFS (mean follow-up: 26.4 months), offering broader prognostic applicability. An important consideration for clinical translation is model interpretability and workflow integration. Radiologists are more likely to adopt AI tools that provide transparent, explainable outputs rather than black-box predictions. In our framework, segmentation masks offer intuitive visualizations of tumor boundaries, while feature-based classification can be complemented with attention heatmaps or feature importance scores to highlight imaging regions or descriptors driving the decision. From a workflow perspective, the system could be integrated into existing ultrasound workstations or PACS environments, where, after the radiologist acquires images, the model automatically generates segmentation, subtype classification, and estimated PFS stratification within seconds. The outputs could be displayed as an overlay on the original scan, accompanied by probability scores and a structured report. Such an approach positions the tool as a decision-support system, assisting rather than replacing radiologists, thereby enhancing diagnostic confidence, reducing inter-observer variability, and facilitating personalized treatment planning. Accurate subtyping of ovarian tumors from ultrasound can directly inform clinical decision-making, particularly in differentiating benign from borderline and malignant subtypes where management strategies diverge significantly. Furthermore, the ability to predict PFS from preoperative imaging provides valuable prognostic insight, potentially guiding personalized treatment planning, risk stratification, and follow-up intensity. By integrating both diagnostic and prognostic tasks into a unified pipeline, our framework has the potential to reduce diagnostic delays, decrease observer variability, and support precision oncology in ovarian cancer care.

In summary, this multicenter deep learning framework demonstrates that integrating radiomic descriptors with transformer- and CNN-derived features enables robust and generalizable ovarian tumor analysis across diverse ultrasound platforms. The results highlight the diagnostic value of combining segmentation, multi-class classification, and PFS prediction into a unified model, while reducing observer variability in ultrasound interpretation. Nevertheless, limitations include the retrospective design, potential center-specific biases despite ICC-based harmonization, and the absence of prospective validation. Additionally, external test performance suggests the need for improved cross-domain adaptation to address scanner and population heterogeneity. Future efforts will focus on prospective clinical trials, multi-modal integration, and explainability to support clinical adoption. While ViT-based models demonstrated superior performance, their reliance on large datasets and high computational demands may limit their clinical deployment, especially in resource-constrained settings. Addressing these challenges through model optimization, hybrid CNN–Transformer architectures, or knowledge distillation techniques will be essential to ensure scalability, efficiency, and broader real-world impact.

Ovarian tumors are a leading cause of cancer-related mortality in women, and early, accurate diagnosis is critical for improving patient outcomes [ 1 ]. However, because symptoms are often absent at early stages, diagnosis typically occurs late, complicating treatment [ 2 ]. Medical imaging is central to detection and characterization, with ultrasound being particularly valuable due to its non-invasive nature, real-time feedback, and accessibility [ 3 ]. Ovarian tumors present a major diagnostic challenge, as ultrasound interpretation is highly operator-dependent and subject to variability across institutions [ 4 ]. Reliable and reproducible diagnostic tools are urgently needed to improve early detection, risk stratification, and treatment planning [ 5 , 6 ]. Accurate segmentation of ovarian tumors in ultrasound remains challenging due to variations in tumor morphology and the presence of noise and artifacts [ 7 , 8 ]. Traditional image processing methods have faced limitations in handling these variations, especially when distinguishing benign from malignant lesions. Recent advances in deep learning [ 9 ] have shown promise, with models such as Unet [ 10 ], nnU-Net [ 11 ], UNETR [ 12 ], Swin-Unet [ 13 ], and SegNet [ 14 ] achieving strong performance in segmentation. While convolutional neural networks (CNNs) [ 15 ] excel at pixel-level analysis, they may miss higher-order structural and statistical information crucial for tumor characterization. Radiomic features, capturing quantitative descriptors of texture and intensity [ 16 ], can complement deep features [ 17 , 18 ] derived from architectures such as ResNet [ 19 ] and Vision Transformers (ViT) [ 20 – 22 ], enhancing model robustness. Handcrafted radiomic features and deep learning representations provide complementary perspectives on tumor biology [ 23 – 25 ]. Radiomics captures interpretable, handcrafted descriptors of tumor intensity, texture, and morphology, which have shown prognostic value in ovarian and other cancers. In contrast, deep learning models, particularly transformer-based architectures, can automatically learn high-level spatial and contextual patterns from imaging data, offering strong generalization across datasets [ 26 , 27 ]. Recent applications of ViTs in medical imaging have demonstrated promising results for tasks such as lesion detection, organ segmentation, and disease classification [ 22 , 28 – 30 ]. By combining these two modalities, our framework leverages both the interpretability of radiomics and the representational power of ViTs to achieve a more comprehensive tumor characterization. Prior computational approaches have largely focused on either handcrafted radiomics or deep learning-based models [ 31 – 33 ], but these methods often struggle with reproducibility and generalization. Variability in ultrasound acquisition protocols, scanner vendors, and patient populations across centers can significantly affect feature stability and model performance. Recent multicenter developments aiming to integrate segmentation with outcome prediction include the Adaptive Multi-modality Segmentation-to-Survival model (AdaMSS), which performs PET/CT tumor segmentation and PFS prediction across nine centers [ 34 ]; and a fully automated MRI-based pipeline for extramural vascular invasion (EMVI) classification and response prediction, validated across multiple institutions [ 35 ]. Additionally, a study combining ultrasound and MRI automatic segmentation to derive a multi-modal radiomics signature demonstrated promising performance for DFS prediction in breast cancer across training and external test sets [ 36 ]. These works underscore the growing trend toward comprehensive imaging pipelines that simultaneously support segmentation and prognostic tasks. While prior studies have typically addressed segmentation, classification, or survival prediction separately [ 4 , 7 , 37 , 38 ], our work is novel in unifying all three tasks—tumor segmentation, cancer classification, and PFS estimation—within a single multicenter deep learning framework. By jointly optimizing these interrelated tasks across diverse institutional datasets, our approach aims to enhance generalizability and clinical utility in oncologic imaging. This study proposes a novel multicenter multi-task learning framework for ovarian tumor analysis that combines segmentation, multi-class classification, and prognostic prediction. The fusion of handcrafted radiomic biomarkers with deep learning embeddings, coupled with an emphasis on reproducibility across centers, enables the framework to deliver accurate and generalizable results. External validation on large multicenter datasets further supports its reliability for real-world deployment. The main innovations of this study are: (i) a unified multi-task approach integrating segmentation, classification, and prognosis; (ii) combining radiomic biomarkers with deep representations from ResNet and ViT for a holistic feature profile; (iii) reproducibility-driven feature selection using ICC; and (iv) multicenter validation to maximize generalizability and clinical relevance.

Additional file 1. Additional file 1.