2.5D ABUS deep learning model with BI-RADS shape and margin auxiliary tasks achieves AUC 0.91 for breast lesion classification

A 2.5D ABUS deep learning model with BI-RADS shape and margin auxiliary tasks achieved an external test AUC of 0.91 (95% CI 0.83-0.96) for breast lesion classification, outperforming 3D Swin Transformer (0.76) and other benchmarks.

• Retrospective study included 387 breast lesions (106 malignant) from 313 patients across two centers; external test set of 72 lesions (21 malignant).
• Model used Swin Transformer V2-T backbone with ResMask fusion for shape/margin features and auxiliary BI-RADS-aligned tasks.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.