Study reveals why mammography AI loses fine malignancy cues

A new study from researchers Vinceline Bertrand and Ionut Cardei uses gradient-based orthogonal latent decomposition for hierarchical Variational Autoencoders (H-VAEs) to mechanistically explain a persistent asymmetry in weakly supervised mammography models: coarse lesion-type features are preserved, but fine-grained malignancy cues degrade. The latent space is partitioned into a task-aligned component (z_1) shaped by coarse supervision and an orthogonal residual (z_res) capturing remaining capacity. On 3,550 mammographic ROIs from CBIS-DDSM, only ~4.4% of latent magnitude aligns with supervisory gradients, leaving ~95.6% in z_res, upon which fine-grained pathology prediction primarily depends.

The model achieves Stage-1 AUC 0.866 and Stage-2 AUC 0.552, with a reconstruction stability gap (Δ_diag=5%, p=0.005) and a classification gap (Δ_AUC=0.314, p<0.001). Latent ablation confirms features for both tasks reside heavily in z_res, explaining why reconstruction degrades pathology stability disproportionately. Comparisons with MIL and MTL confirm generalization across architectures. The findings reveal that single coarse supervision isolates only a sparse 1D latent direction, forcing critical fine-grained features into a vulnerable residual subspace—a structural flaw that directly impacts clinical reliability of breast cancer screening pipelines.