Abstract
In mammographic interpretation, lesion type is typically characterized first, but the malignancy assessment ultimately drives recall and biopsy decisions—and this decision may be less protected under weak supervision. We introduce gradient-based orthogonal latent decomposition for hierarchical Variational Autoencoders (H-VAEs) to investigate this asymmetry. The method partitions the latent space into a taskaligned component (z1), shaped by coarse supervisory gradients, and an orthogonal residual component (zres). Across 3,550 mammographic regions of interest from CBIS-DDSM, only ∼4.4% of the latent magnitude aligns with the supervisory gradients, leaving ∼95.6% in the residual space on which both tasks depend. Reconstruction also perturbs the pathology head more strongly than the lesion-type head (Δdiag = 5.0%, p = 0.005). Because the pathology head is only weakly discriminative (AUC 0.552, 95% CI 0.500–0.602), we interpret the mammography findings as suggestive rather than conclusive. However, replication on chest X-ray, where the fine-grained head performs well above chance (AUC 0.820), reveals the same asymmetry. These results provide breast imaging practitioners with a representation-level diagnostic that aggregate AUC alone does not capture.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/Deep_Brea3th_003.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=7UUi8ma4G4
BibTex
@InProceedings{BerVin_GradientBased_MICCAISAT2026,
author = { Bertrand, Vinceline AND Cardei, Ionut},
title = { { Gradient-Based Latent Decomposition Reveals Mechanisms of Feature Degradation in Weakly Supervised Mammography } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17256},
month = {pending},
page = {pending}
}
