Related Experiment Video
Updated: Aug 11, 2026
10:21
A Coregistered Ultrasound and Photoacoustic Imaging Protocol for the Transvaginal Imaging of Ovarian Lesions
Published on: March 3, 2023
From ACR O-RADS 2022 to Explainable Deep Learning: Comparative Performance of Expert Radiologists, Convolutional
Ali Abbasian Ardakani1, Afshin Mohammadi2, Alisa Mohebbi1
1Department of Medicine, Faculty of Medicine and Dentistry, Danube Private University, Krems an der Donau, Austria.
Summary
Deep learning models significantly outperform radiologist assessments for ovarian lesions using O-RADS v2022. Combining radiologist scores with AI models offers the highest accuracy for diagnosing ovarian masses.
Area of Science:
- Medical imaging and artificial intelligence
- Ovarian lesion characterization
- Diagnostic accuracy in radiology
Background:
- The Ovarian-Adnexal Reporting and Data System (O-RADS) version 2022 enhances adnexal lesion risk stratification.
- Radiologist interpretation of O-RADS scores shows inter-observer variability and can be overly conservative.
- Deep learning (DL) models show potential for improving ovarian mass characterization.
Purpose of the Study:
- To evaluate radiologist performance using O-RADS v2022 for adnexal lesions.
- To compare the diagnostic performance of convolutional neural network (CNN) and vision transformer (ViT) models against radiologists.
- To investigate the diagnostic benefits of hybrid human-AI frameworks, emphasizing explainable AI.
Main Methods:
- Retrospective analysis of 512 ultrasound images from 227 patients, including 110 with malignant lesions.
- Training and validation of 16 DL models (DenseNets, EfficientNets, ResNets, VGGs, Xception, ViTs).
- Development of hybrid frameworks integrating radiologist O-RADS scores with DL-predicted malignancy probabilities.
Main Results:
- Radiologist-only O-RADS v2022 assessment achieved an AUC of 0.683 and 68.0% accuracy.
- DL models showed varying performance, with ViT16-384 achieving the highest AUC of 0.941 and 87.4% accuracy.
- Hybrid human-AI frameworks significantly improved the performance of most CNNs and ViTs (p < .05).
Conclusions:
- Deep learning models demonstrate superior performance compared to radiologist-only O-RADS v2022 assessments.
- Hybrid human-AI approaches, integrating radiologist expertise with AI predictions, achieve the highest diagnostic accuracy.
- These hybrid paradigms hold promise for standardizing ultrasound interpretation, reducing false positives, and improving the detection of high-risk ovarian lesions.
