Characterize Pathology Encoding in Unsupervised Renal Representations

Characterize how pathological structures are encoded within representations learned without abnormality supervision, determine which network layers most strongly express abnormality-related information, and establish whether different pathological processes produce distinct changes in those representations.

Background

The paper uses a frozen residual U-Net backbone from Omni-Seg, which was pretrained to segment structurally normal renal primitives without abnormality-specific supervision. NoRDeC combines Mahalanobis-distance anomaly scoring with layer-wise representation analysis using centered kernel alignment (CKA). The authors report that intermediate-to-deep encoder features support abnormality detection and that different glomerular abnormality categories exhibit different residual CKA patterns relative to institution-specific normal references.

Despite these findings, the paper identifies an unresolved representational question: anomaly scores indicate whether a glomerulus deviates from normal morphology but do not explain how pathological information is distributed through the network hierarchy or whether the observed representation changes are specific to distinct pathological processes. Resolving this issue would improve interpretation of normal-reference anomaly detectors and clarify the biological and computational meaning of their learned features.

References

Although these methods have demonstrated strong anomaly-detection and localization performance, they are designed primarily to produce anomaly scores and localization maps. They do not explicitly characterize how abnormality-related information changes across the network hierarchy. Therefore, it remains unclear how pathological structures are encoded within representations learned without abnormality supervision, at which layers abnormality-related information is most strongly expressed, and whether different pathological processes produce distinct changes in these representations.

— Seeing Abnormal from Normal: Glomerular Abnormality in Representations of Normal Renal Morphology  (2609.19444 - Hasko et al., 16 Sep 2026) in Section 1, Description of purpose

Neither traces a finding through the training objective, so whether the objective itself removes local pathology remains open.

— The Predictability Bias of Joint-Embedding Predictive Learning in Medical Imaging  (2610.13237 - Drexel et al., 9 Oct 2026) in Section 2, Related Work, p. 2