ABUS-ResMask-Net: ABUS Lesion Classification Network Design with ResMask Module and BI-RADS Content-Awareness Auxiliary Tasks.

This study aims to develop a 2.5D deep learning model with shape and margin as auxiliary tasks to improve the diagnostic performance of benign-malignant classification of breast lesions in automated breast ultrasound system (ABUS) images. In this retrospective study, 387 breast lesions (106 malignant and 281 benign) from 313 patients were enrolled from two centers between 2021 and 2024. Lesions from Center 1 (315 lesions, 85 malignant, 230 benign) were used as the training cohort, and lesions from Center 2 (72 lesions, 21 malignant, 51 benign) were used as the held-out external testing cohort, with no patient overlap between the two centers. A deep learning classification algorithm, combined with a ResMask Fusion module, was used to extract and integrate morphological features, particularly shape and margin, from two-dimensional ultrasound images. ABUS-ResMask-Net uses Swin Transformer V2-T as the backbone, with shape and margin classification as auxiliary tasks. In the final lesion-level evaluation, ABUS-ResMask-Net achieved an AUC of 0.91 (95% CI: 0.83-0.96). In the benchmark comparison with representative methods, ABUS-ResMask-Net achieved the highest AUC, compared with 0.76 (95% CI, 0.62-0.87) for 3D Swin Transformer, 0.85 (95% CI, 0.76-0.93) for Yang et al., and 0.87 (95% CI, 0.79-0.94) for BI-RADS-Net-v2. The proposed 2.5D ABUS-ResMask-Net, which combines BI-RADS-aligned auxiliary tasks with the ResMask Fusion module, achieved higher lesion-level AUC than representative 3D, 2.5D, and BI-RADS-oriented comparison methods on the same external testing cohort, suggesting its potential for ABUS-based benign-malignant lesion classification.