Gradient-Based Latent Decomposition Reveals Mechanisms of Feature Degradation in Weakly Supervised Mammography
Gradient-Based Latent Decomposition Reveals Mechanisms of Feature Degradation in Weakly Supervised Mammography
基于梯度潜空间分解揭示弱监督乳腺X光摄影中特征退化的机制
Abstract: Weakly supervised hierarchical models exhibit a persistent asymmetry: coarse lesion-type features are preserved under reconstruction while fine-grained malignancy cues degrade—a pattern with direct consequences for the clinical reliability of breast cancer screening pipelines.
摘要： 弱监督分层模型表现出一种持续的不对称性：粗略的病变类型特征在重建过程中得以保留，而细粒度的恶性肿瘤线索却发生退化——这种模式直接影响了乳腺癌筛查流程的临床可靠性。
We introduce gradient-based orthogonal latent decomposition for hierarchical Variational Autoencoders (H-VAEs) to mechanistically explain this asymmetry. The latent space is partitioned into a task-aligned component ($z_1$), shaped by coarse supervisory gradients, and an orthogonal residual ($z_{\text{res}}$) capturing remaining representational capacity.
我们为分层变分自编码器（H-VAEs）引入了基于梯度的正交潜空间分解，以从机制上解释这种不对称性。潜空间被划分为两个部分：由粗略监督梯度塑造的任务对齐分量（$z_1$），以及捕获剩余表征能力的残差分量（$z_{\text{res}}$）。
On 3,550 mammographic Regions of Interest (ROIs) from CBIS-DDSM, only ~4.4% of latent magnitude aligns with supervisory gradients, leaving ~95.6% in the orthogonal residual 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 of $\Delta_{\text{diag}}=5%$ ($p=0.005$) and a classification gap of $\Delta_{\text{AUC}}=0.314$ ($p<0.001$).
在来自 CBIS-DDSM 的 3,550 个乳腺X光摄影感兴趣区域（ROI）上，仅约 4.4% 的潜空间量级与监督梯度对齐，而约 95.6% 的量级留在了正交残差中，细粒度的病理预测主要依赖于此。该模型在第一阶段的 AUC 为 0.866，第二阶段的 AUC 为 0.552，重建稳定性差距为 $\Delta_{\text{diag}}=5%$ ($p=0.005$)，分类差距为 $\Delta_{\text{AUC}}=0.314$ ($p<0.001$)。
Latent ablation confirms that features for both tasks reside heavily in $z_{\text{res}}$, structurally explaining why reconstruction degrades pathology stability disproportionately. Comparisons with Multi-Instance Learning (MIL) and Multi-Task Learning (MTL) confirm generalization across architectures and modalities.
潜空间消融实验证实，两项任务的特征都大量存在于 $z_{\text{res}}$ 中，这从结构上解释了为什么重建过程会不成比例地降低病理预测的稳定性。与多示例学习（MIL）和多任务学习（MTL）的对比证实了该方法在不同架构和模态下的泛化能力。
These findings reveal that in high-dimensional spaces, a single coarse supervisory signal isolates only a sparse 1D latent direction, forcing critical fine-grained features into the vulnerable residual subspace.
这些发现揭示了在高维空间中，单一的粗略监督信号只能分离出一个稀疏的一维潜方向，从而迫使关键的细粒度特征进入脆弱的残差子空间。