Lesion-aware hybrid CNN-transformer improves mammography classification on VinDrMammo

A hybrid convolutional neural network (CNN)-transformer mammography model reached 85.5% accuracy and 79.6% macro F1 on VinDrMammo three-class classification, with lesion-focused evidence matching annotated abnormalities and fewer parameters than comparators.

• On the VinDrMammo three-class task, LENS-Base outperformed ConvNeXt, DINOv2, GMIC, and Swin Transformer under a unified protocol.
• The weakly supervised lesion-aware branch produced regional evidence that frequently overlapped annotated abnormalities, supporting interpretability.
• LENS achieved these results with fewer parameters and lower FLOPs than advanced architectures.

Automated summary

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