VGG11-based deep learning for automated MRI classification of adenomyosis: a systematic comparison of five classification criteria

Magnetic resonance imaging (MRI) has been increasingly applied in the classification of adenomyosis; however, due to the lack of a standardized gold standard for lesion classification, it remains challenging to make an early diagnosis and initiate timely treatment. This study aimed to develop a deep learning (DL)-based classification system for adenomyosis and systematically compare the performance of five most frequently researched classification criteria. Using standardized protocols (slice thickness: 3–4 mm, TR/TE: 3000–5000/80–100 ms), T2-weighted images (T2WI) from 6,768 MRI scans acquired at 1.5T and 3T scanners (Siemens, GE, and Philips) were analyzed. Images were independently labeled by three expert radiologists with at least 10 years of experience according to five classification systems (Hulka, Kishi, Bazot, Kobayashi, and Gong), with interobserver agreement (κ = 0.82–0.91). The VGG11 model was evaluated in terms of accuracy, precision, recall, and F1-score with 95% confidence intervals (CIs). The model achieved over 90% accuracy on all criteria. The Bazot 3-class system showed the best performance (accuracy: 0.9778 [95% CI: 0.972–0.983], F1-score: 0.9723). K-fold cross-validation showed stable convergence within 200 epochs. Confusion matrix analysis revealed an exceptional generalization capability, even for the extreme class imbalances (at the sample ratio of 86:1). The Gong 6-class system achieved lower performance (accuracy: 0.9214 [95% CI: 0.912–0.931]) due to data imbalance. Our DL system provided accurate, automated adenomyosis classification. The systematic comparison revealed significant performance differences among various criteria, and the Bazot system was optimal for DL applications, which endowed clinicians with the enhanced diagnostic capabilities for adenomyosis evaluation through MRI. The VGG11-based deep learning model achieved exceptional diagnostic accuracy (over 90% on all criteria), establishing a new benchmark for the MRI-based detection of adenomyosis. This study provided the first comprehensive evaluation of the five most frequently researched adenomyosis classification systems using deep learning, revealing significant performance differences that inform the selection of the optimal clinical protocol. The demonstrated accuracy and consistency of our automated system addressed critical standardization gaps in adenomyosis diagnosis, providing an objective tool for clinicians to reduce diagnostic variability and improve treatment planning.