Source-linked AI summary
Less Contouring, More Accuracy: Lesion-Guided ROI Deep Learning for Ovarian Ultrasound Classification
Mehran Ahmad, Ali Abbasian Ardakani, Afshin Mohammadi, Alisa Mohebbi, Gernot Kronreif, Sepideh Hatamikia
TL;DR
Ovarian ultrasound classification is challenged by overlapping lesion appearances and operator-dependent interpretation. The study compares four AI strategies across multiclass and binary public datasets, finding that lesion-guided ROI deep learning offers the strongest balance of performance and annotation burden. Prospective validation on heterogeneous clinical data and additional imaging modalities remains necessary before implementation.
Problem
Overlapping sonographic features and operator dependence make objective ovarian lesion classification challenging and motivate comparison of less annotation-intensive lesion-focused strategies.
Method
Four classification strategies were evaluated under a unified framework across the eight-class MMOTU dataset and binary DF-versus-PCO OUD task, using deep learning architectures and contour-based radiomics with machine learning.
Results
Lesion-guided ROI achieved the strongest overall performance; on OUD, lesion-guided MaxViT-Tiny reached 97.56% accuracy, while contour-based ResNet18 matched lesion-guided ResNet18 at 92.68%.
Takeaways & Limitations
Lesion-guided ROI localization may provide a practical, scalable balance between diagnostic performance and annotation feasibility for ovarian ultrasound classification.
Takeaways & Limitations
The study used public B-mode ultrasound datasets; further validation with heterogeneous clinical data, additional modalities, and computational evaluation is needed before clinical implementation.
Abstract
from arXiv · showhide
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
1 Introduction
Transvaginal ultrasound supports ovarian lesion evaluation but remains operator-dependent because benign and malignant lesions can share overlapping or subtle sonographic features. This study compares AI strategies, including lesion-guided localization, to improve objective classification while reducing annotation demands.
- Transvaginal sonography provides real-time, high-resolution visualization for ovarian mass evaluation and supports risk stratification through IOTA and O-RADS systems.
- Overlapping sonographic features and subtle distinctions among lesion etiologies contribute to substantial inter-observer variability in ultrasound interpretation.
- Deep learning and radiomics models can assist ovarian lesion characterization by providing rapid probability estimates or quantitative imaging biomarkers.
- Prior AI studies reported promising performance in binary and multiclass ovarian ultrasound classification, including AUC values of 0.911 and 0.93.
- The study compares global image-based DL, lesion-guided ROI-based DL, lesion contour-based DL, and contour-based radiomics across binary and multiclass settings.
2 Materials and methods
The study compares four ovarian ultrasound classification strategies across two publicly available datasets under a unified experimental framework. It emphasizes lesion-guided ROI learning as a way to retain local context while reducing reliance on pixel-level contour annotation.
- Datasets: The MMOTU dataset was evaluated as an eight-class task, whereas OUD was reformulated as binary DF-versus-PCO classification because of limited normal-ovary data.
- Classification strategies: Four strategies used global images, lesion-guided rectangular ROIs, expert-annotated lesion contours, or contour-isolated radiomics features.
- Classification strategies: Lesion-guided ROI learning retained the lesion and immediate local context while reducing irrelevant background and the annotation demands of precise contouring.
- Radiomics: Contour-based radiomics extracted 215 features, comprising 102 PyRadiomics descriptors and 113 handcrafted ultrasound-specific features, with ANOVA-based selection applied for OUD.
- Experimental protocol: All strategies used consistent dataset splits and test-set evaluation, while deep learning models included MaxViT-Tiny, Swin Transformer, EfficientNet-B7, and ResNet-18.
3 Results
Lesion-focused deep learning achieved the strongest overall classification performance across the MMOTU and OUD datasets, while radiomics performed less well on the multiclass task. Lesion-guided ROI cropping matched contour-based performance with lower annotation burden and retained useful perilesional context.
- Study overview: 1,729 ultrasound images from 507 patients were analyzed across MMOTU and OUD.MMOTU contributed 1,469 images from 247 patients, and OUD contributed 260 images from 260 patients.
- Global image-based deep learning: 81.6% top-1 accuracy was achieved by MaxViT-Tiny with global images on the eight-class MMOTU dataset.Global-image models achieved 71.1%–81.6% top-1 accuracy and AUC values of 0.93–0.95 on MMOTU.
- Global image-based deep learning: 95.1% top-1 accuracy was achieved by MaxViT-Tiny with global images on the binary OUD dataset.Global-image models achieved 85.4%–95.1% top-1 accuracy and AUC values of 0.87–0.99, although confidence intervals overlapped considerably.
- Lesion-guided ROI-based deep learning: 93.1% top-1 accuracy on MMOTU and 97.6% on OUD were achieved by MaxViT-Tiny using lesion-guided ROI crops.ROI models achieved AUC values of 0.98–0.99 on MMOTU and 0.97–0.99 on OUD, with OUD top-2 accuracy uniformly at 100%.
- Lesion contour-based deep learning: 91.8% MMOTU top-1 accuracy and 95.1% OUD top-1 accuracy were achieved by the strongest contour-based deep learning models.Contour-based performance was comparable to lesion-guided ROI models, but complete removal of perilesional context may discard supportive features.
- Lesion contour-based radiomics: 74.2% top-1 accuracy was the best radiomics result on MMOTU, whereas all evaluated OUD radiomics classifiers reached 95.12% top-1 accuracy with 60 selected features.OUD top-2 accuracy reached 100%; the multiclass performance gap suggests fixed quantitative descriptors may not capture complex eight-class patterns sufficiently.
- Overall comparison: Lesion-guided ROI cropping provided the most practical balance between diagnostic performance and annotation burden.It preserved limited adjacent structures while requiring less detailed annotation than contour-based approaches.
4 Discussion
Across the evaluated ovarian ultrasound strategies, lesion-guided ROI learning offered the strongest practical balance between classification performance and annotation burden. The comparisons also identify lesion representation as an important consideration, while broader validation remains necessary before clinical implementation.
- Lesion-guided ROI-based learning provided the most favorable balance between classification performance and annotation burden.It achieved the strongest overall performance while avoiding the detailed annotation required for contour-based approaches.
- A modest perilesional margin retains lesion–tissue interface cues while excluding distant background and acquisition artifacts.This provides a stated rationale for why lesion-centered crops can preserve relevant tumor information without precise pixel-level contouring.
- Explicit ROI guidance may improve capture of textural and structural distinctions among lesions with overlapping appearances.The discussion links this to potentially more consistent outputs and support for less experienced operators in triage, follow-up, and referral.
- 93.10% top-1 accuracy was achieved by lesion-guided MaxViT-Tiny on the eight-class MMOTU task.This exceeded the previously reported MMOTU accuracies cited for studies and.
- 97.56% accuracy was achieved by lesion-guided MaxViT-Tiny on the binary OUD task, while contour-based ResNet18 matched lesion-guided ResNet18 at 92.68%.Under the within-study binary setting, lesion guidance improved global ResNet18 accuracy from 85.37% to 92.68%, a 7.31% improvement.
- Further validation using heterogeneous clinical data, additional imaging modalities, and computational-resource evaluation is needed before clinical implementation.The authors also call for prospective multicenter validation across clinical settings, patient populations, and ultrasound systems.
5 Conclusion
The study concludes that lesion-guided ROI deep learning may balance diagnostic performance with annotation feasibility in ovarian ultrasound classification. Prospective multicenter validation is needed to confirm performance.
- Lesion-guided ROI deep learning may offer a practical balance between diagnostic performance and annotation feasibility.The approach directs attention to the lesion while retaining limited surrounding context.
- Prospective multicenter validation is needed to confirm the proposed framework's performance.
- The approach could support more consistent interpretation with reduced clinician burden compared with global or strictly contour-based methods.
Declarations
The study reports its funding, competing-interest status, data provenance, and use of publicly available de-identified datasets.
- The research was funded by the FTI-Dissertationen grant, project number FTI24-D-008.
- The authors declare no competing interests.
- The study used secondary analysis of publicly available, de-identified ovarian ultrasound datasets without recruiting participants or collecting new patient data.Ethical approval and consent for original data collection were reported by the dataset providers.
- The MMOTU and OUD datasets analyzed in the study are publicly available.