A lesion-guided artificial-intelligence approach recorded high accuracy in classifying ovarian ultrasound images while demanding less detailed marking of lesions than methods based on pixel-by-pixel contours, according to a new preprint. The strongest reported results came from a model called MaxViT-Tiny, which reached 93.10% accuracy on MMOTU and 97.56% on OUD. Both results also had a reported AUC of 0.99.
The analysis covered 1,729 ultrasound images from 507 patients: 1,469 images from 247 patients in MMOTU and 260 images from 260 patients in OUD.
The researchers compared four ways of presenting images to an AI system: the full scan, a lesion-guided region of interest, an expert-drawn lesion contour, and radiomics extracted from the contour. Radiomics is a numerical-feature approach for analysing image patterns. The deep-learning comparisons used MaxViT-Tiny, Swin Transformer, EfficientNet-B7 and ResNet18, while the radiomics analysis used support-vector machines, nearest-neighbour classification and artificial neural networks.
What the comparison tested
MMOTU supported a multiclass classification task. OUD was treated differently: the researchers reduced it to a binary comparison between dominant-follicle and polycystic-ovary images after excluding its limited normal-ovary class. That means the OUD results are not a direct test of the original three-class problem.
To make the strategies easier to compare, all were evaluated with the same 70:15:15 split for training, validation and testing. Test-set uncertainty was estimated through 1,000 bootstrap resamples, producing percentile-based 95% confidence intervals for the reported metrics.
For the OUD radiomics work, one-way analysis of variance ranked 215 candidate image features. Feature selection and preprocessing were fitted only on the training partition, rather than using information from the held-out data.
The trade-off between detail and performance
The contour-based deep-learning systems performed at a similar level in the reported comparison, reaching 91.81% accuracy with an AUC of 0.99 on MMOTU and 95.12% accuracy with an AUC of 0.99 on OUD. But the authors said contouring required substantially more annotation effort. The study did not provide a numerical estimate of the time or labour involved.
An architecture-matched comparison on OUD offered another indication of the difference between broad and focused inputs. ResNet18 scored 85.37% accuracy when given the global image and 92.68% when given lesion-guided ROI inputs, reported as a 7.31% improvement. Because this was a retrospective dataset analysis, the comparison does not establish that ROI marking itself caused the higher score.
The contour-based radiomics results differed across the two tasks. On MMOTU, top-1 accuracy ranged from 71.55% to 74.2%, with the radial-basis-function support-vector machine producing the best reported result. On OUD, retaining 60 selected features produced 95.12% top-1 accuracy for the support-vector machine, nearest-neighbour and neural-network classifiers, while top-2 accuracy reached 100%.
The authors interpret lesion-guided deep learning as the most practical balance between classification performance and the effort needed to locate a lesion. They suggest it could support more consistent interpretation.
Why the results are not yet a clinical test
One important limitation affects the MMOTU evaluation. Its 469-image held-out portion was divided image by image into 237 validation images and 232 final-test images because patient identifiers were unavailable. As a result, the analysis could not verify patient-level independence between those partitions.
The study used only public B-mode ultrasound datasets. The authors said that heterogeneous clinical data, additional ultrasound modalities and the computational requirements of deployment would need to be evaluated before clinical implementation. No prospective multicentre or external clinical testing was performed in the analysis.
The manuscript is an arXiv preprint, identified as arXiv:2608.25965v1 and dated 26 August 2026.
Funding and access
The research was funded by an FTI-Dissertationen grant, project FTI24-D-008. The authors reported no competing interests. The work was a secondary analysis with no new recruitment or data collection, and ethics and consent for the original data collection were reported by the dataset providers.
The MMOTU and OUD datasets analyzed in the study are publicly available, and supplementary material is stated to be available with the online version of the article.
Paper data and sources
Original title: Less Contouring, More Accuracy: Lesion-Guided ROI Deep Learning for Ovarian Ultrasound Classification
Authors: Mehran Ahmad, Ali Abbasian Ardakani, Afshin Mohammadi et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text