A new AI framework for breast cancer classification gets better results by making two mammogram views talk to each other more carefully.
Most mammography screening uses two angles — craniocaudal (CC, top-down) and mediolateral oblique (MLO, side-angled) — because neither view alone tells the full story. Existing multi-view AI models tend to merge these views crudely, either blending features in one shot or running a single round of cross-attention. A team of researchers proposes a different approach: dedicated "fusion tokens" that shuttle information between the two views at multiple layers inside a frozen vision transformer, rather than at just one point. The backbone model stays locked; only the fusion tokens and lightweight prompt layers are trained.
The distinction matters because where fusion happens inside a neural network affects what the model actually learns. Pushing cross-view communication through multiple transformer depths lets the network build progressively richer representations of abnormalities that appear differently — or only partially — in each angle. On the VinDr-Mammo dataset, the framework reached a 0.8090 AUC and 50.40% F1 on BI-RADS classification, including a 0.10 AUC gain over a standard dual-view fusion baseline in the binary setting. Results on the CMMD dataset also improved over simpler baselines.
Breast cancer AI has attracted serious research attention for years, but mammography remains a hard problem: class imbalance, subtle lesions, and variation in imaging equipment all erode model performance. A 0.10 AUC bump is meaningful, though the absolute F1 of 50.40% is a reminder that no one is replacing a radiologist yet.