Organizations: Department of Medicine, Danube Private University (DPU), Krems, Austria. · Austrian Centre for Medical Innovation and Technology (ACMIT), Wiener Neustadt, Austria. · Department of Medical Physics and Biomedical Engineering, Medical University of Vienna, Vienna, Austria. · Department of Radiology, Faculty of Medicine, Urmia University of Medical Science, Urmia, Iran.
Ovarian lesion classification using transvaginal ultrasound remains challenging due to overlapping imaging characteristics and the dependence on expert interpretation. This study investigates whether lesion-guided region-of-interest (ROI) deep learning can achieve competitive diagnostic performance while reducing the annotation burden associated with pixel-level lesion segmentation. Two publicly available ovarian ultrasound datasets were evaluated: the Multi-Modality Ovarian Tumor Ultrasound (MMOTU) dataset for eight-class classification and the Ovarian Ultrasound Dataset (OUD) for binary classification. Four strategies were compared under a unified framework: global image-based deep learning, lesion-guided ROI-based deep learning, lesion contour-based deep learning, and contour-based radiomics with machine learning classifiers. Four deep learning architectures, MaxViT-Tiny, Swin Transformer, EfficientNet-B7, and ResNet18, were evaluated. Radiomics models were developed using support vector machine, k-nearest neighbors, and artificial neural network classifiers, with ANOVA-based feature selection applied for the lower-sample OUD dataset. The lesion-guided ROI strategy achieved the strongest overall performance, with MaxViT-Tiny obtaining 93.10% accuracy and an AUC of 0.99 on MMOTU and 97.56% accuracy and an AUC of 0.99 on OUD. The contour-based approach achieved comparable accuracy but required substantially higher annotation effort. These findings demonstrate that lesion-guided ROI deep learning provides an effective balance between diagnostic performance and annotation efficiency, offering a practical approach for scalable AI-assisted ovarian ultrasound analysis
Figures & tables
Figure 1: Clinical challenges in visual ovarian ultrasound interpretation and the role of AI-assisted diagnosis.
Figure 2: The three deep learning strategies are based on the global ultrasound image, lesion-guided ROI, and expert-annotated lesion contour.
Figure 3: The lesion contour-based radiomics strategy, including lesion masking, extraction of PyRadiomics and handcrafted ultrasound features, feature integration, and classification using SVM, KNN, and ANN.
Data
Model
Top-1 Accuracy (%)
Top-2 Accuracy (%)
Specificity (%)
Sensitivity (%)
AUC
MMOTU
MaxViT-Tiny
81.57 [76.72, 85.78]
90.09 [86.21, 93.53]
96.69 [90.80, 100.00]
81.57 [76.72, 85.78]
0.94 [0.92, 0.96]
Swin
78.45 [73.28, 83.62]
91.00 [86.21, 93.55]
96.22 [89.23, 100.00]
78.45 [73.28, 83.62]
0.95 [0.9386, 0.97]
EfficientNet-B7
79.31 [74.14, 84.48]
89.66 [85.34, 93.10]
96.24 [89.48, 100.00]
79.31 [74.14, 84.48]
0.95 [0.93, 0.97]
ResNet18
71.12 [65.09, 76.72]
85.78 [80.60, 90.09]
95.68 [88.43, 99.01]
71.12 [65.09, 76.72]
0.93 [0.91, 0.95]
OUD
MaxViT-Tiny
95.12 [87.80, 100.00]
100.00 [100.00, 100.00]
94.97 [86.80, 100.00]
95.12 [87.80, 100.00]
0.99 [0.96, 1.00]
EfficientNet-B7
90.24 [80.49, 97.56]
100.00 [100.00, 100.00]
87.54 [78.05, 97.56]
90.24 [80.49, 97.56]
0.99 [0.96, 1.00]
Table 1: Performance of the global image-based deep learning models on the MMOTU and OUD datasets. Values in brackets represent 95% confidence intervals.
Data
Model
Top-1 Accuracy (%)
Top-2 Accuracy (%)
Specificity (%)
Sensitivity (%)
AUC
MMOTU
MaxViT-Tiny
93.10 [89.66, 96.12]
96.98 [94.83, 99.14]
98.99 [94.66, 100.00]
93.10 [89.66, 96.12]
0.99 [0.98, 0.99]
EfficientNet-B7
88.79 [84.48, 92.67]
96.55 [93.97, 98.71]
98.91 [94.51, 100.00]
88.79 [84.48, 92.67]
0.99 [0.98, 0.99]
Swin
87.07 [82.76, 90.95]
98.28 [96.55, 99.57]
97.77 [92.86, 100.00]
87.07 [82.76, 90.95]
0.99 [0.98, 0.99]
ResNet18
85.78 [81.03, 90.10]
93.97 [90.95, 96.56]
98.49 [93.66, 100.00]
85.78 [81.03, 90.10]
0.97 [0.96, 0.98]
OUD
MaxViT-Tiny
97.56 [92.68, 100.00]
100.00 [100.00, 100.00]
98.09 [93.58, 100.00]
97.56 [92.68, 100.00]
0.99 [0.98, 1.00]
EfficientNet-B7
92.68 [82.93, 100.00]
100.00 [100.00, 100.00]
94.28 [90.81, 100.00]
92.68 [82.93, 100.00]
0.97 [0.92, 1.00]
Table 2: Performance of the lesion-guided ROI-based deep learning models on the MMOTU and OUD datasets. Values in brackets represent 95% confidence intervals.
Dataset
Model
Top-1 Accuracy (%)
Top-2 Accuracy (%)
Specificity (%)
Sensitivity (%)
AUC
MMOTU
MaxViT-Tiny
91.81 [88.36, 95.26]
96.98 [94.83, 99.14]
98.97 [95.31, 100.00]
91.81 [88.36, 95.26]
0.99 [0.98, 0.99]
EfficientNet-B7
91.38 [87.50, 94.40]
95.69 [92.67, 98.28]
99.00 [96.01, 100.00]
91.38 [87.50, 94.40]
0.99 [0.9855, 0.99]
Swin
86.21 [81.47, 90.52]
96.98 [94.83, 98.72]
97.87 [94.31, 100.00]
86.21 [81.47, 90.52]
0.99 [0.98, 0.99]
ResNet18
85.78 [81.03, 90.10]
93.97 [90.95, 96.56]
98.49 [95.10, 100.00]
85.78 [81.03, 90.10]
0.98 [0.97, 0.99]
OUD
MaxViT-Tiny
95.12 [87.80, 100.00]
100.00 [100.00, 100.00]
96.18 [89.10, 100.00]
95.12 [87.80, 100.00]
0.99 [0.98, 1.00]
EfficientNet-B7
95.12 [87.80, 100.00]
100.00 [100.00, 100.00]
96.18 [89.10, 100.00]
95.12 [87.80, 100.00]
0.99 [0.97, 1.00]
Table 3: Performance of the lesion contour-based deep learning models on the MMOTU and OUD datasets. Values in brackets represent 95% confidence intervals.
Dataset
Classifier
Top-1 Accuracy (%)
Top-2 Accuracy (%)
Specificity (%)
Sensitivity (%)
AUC
MMOTU
ANN
72.41 [66.81, 77.59]
84.05 [79.31, 88.36]
94.83 [93.51, 96.08]
72.41 [66.81, 77.59]
0.92 [0.90, 0.94]
KNN
71.55 [65.09, 76.72]
84.91 [80.17, 89.23]
94.32 [92.78, 95.60]
71.55 [65.09, 76.72]
0.91 [0.88, 0.93]
SVM-RBF
74.28 [67.67, 78.88]
87.50 [83.19, 91.38]
94.93 [93.80, 96.10]
74.28 [67.67, 78.88]
0.93 [0.91, 0.95]
OUD
ANN
95.12 [87.80, 100.00]
100.00 [100.00, 100.00]
96.18 [89.45, 100.00]
95.12 [87.80, 100.00]
0.98 [0.93, 1.00]
KNN
95.12 [87.80, 100.00]
100.00 [100.00, 100.00]
96.18 [89.45, 100.00]
95.12 [87.80, 100.00]
0.97 [0.92, 1.00]
SVM-RBF
95.12 [87.80, 100.00]
100.00 [100.00, 100.00]
96.18 [89.45, 100.00]
95.12 [87.80, 100.00]
0.95 [0.86, 1.00]
Table 4: Performance of the lesion contour-based radiomics classifiers on the MMOTU and OUD datasets. Values in brackets represent 95% confidence intervals.
Figure 4: Comparison of top-1 accuracies across all evaluated models and classification strategies.
Figure 5: Effect of lesion localization strategy and annotation complexity on ovarian ultrasound classification performance.
Study (Year)
Dataset
Task
Method
Best Accuracy
AUC
(2024) Pham and Le (2024)
MMOTU (public)
8-class
YOLOv8x
70.6% (mAP50)
–
(2025) Zhao et al. (2022)
MMOTU (public)
8-class
EfficientNetV2-M (whole image)
80.60%
–
(2026) Alsubai et al. (2026)
MMOTU (public)
8-class
EfficientOvaNet (EfficientNet-B3, ROI + global; pixel masks)
91.9%
0.980
Proposed
MMOTU (public)
8-class
Lesion-guided ROI + MaxViT-Tiny (single model, bounding box)
93.10%
0.991
(2025) Borna et al. (2025)
OUD (public)
3-class (normal/DF/PCO)
ResNet18
76.2%
–
(2025) Borna et al. (2025)
OUD (Binary)
2-class DF vs PCO
Global image ResNet18 architecture-matched baseline
85.37%
0.879
Table 5: Comparison of the proposed approach with previous studies on the MMOTU and OUD datasets.
SynDiag s.r.l., Turin, Italy · Academic Division of Gynecology and Obstetrics, University of Turin, Azienda Ospedaliera Ordine Mauriziano, Turin, Italy · University of Turin +5
National-Regional Key Technology Engineering Laboratory for Medical Ultrasound, Guangdong Key Laboratory for Biomedical Measurements and Ultrasound Imaging, School of Biomedical Engineering, Shenzhen University Medical School, Shenzhen University, Shenzhen 518055, China · Marshall Laboratory of Biomedical Engineering, Shenzhen University, Shenzhen 518055, China · Department of Ultrasound, Qilu Hospital of Shandong University, Jinan 250012, China +9