Abstract
Frozen medical-image embeddings can support accurate disease probes while also encoding acquisition shortcuts. We audit this structure in DINOv2 embeddings of 35,126 EyePACS-style fundus images using pixel-derived quality variables, mean-difference directions, and targeted residualization. Quartile thresholds, directions, and the removal basis are estimated from training patients only. Diabetic retinopathy (DR) remains linearly decodable (AUC 0.780/0.845/0.938 for any, referable, and severe DR), while five of six bottleneck probes exceed 0.99 AUC. Raw disease–bottleneck cosines reach 0.265, but none exceeds a construction-matched null; the embedding space has effective rank 22 of 768, so raw overlap is descriptive rather than evidence of classifier reliance. Removing a rank-2 sharpness/laterality subspace drives both probes to chance (0.999 to 0.499 and 0.905 to 0.498 on the primary split) with small disease costs (any-DR $-0.0027 \pm 0.0007$ over four patient-level splits). ODIR-5K independently reproduces bottleneck decodability, suppression, and low effective rank. The audit is inexpensive and leakage-free, but it is a diagnostic, not a causal-deconfounding method.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DEMI_022.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=T49wtfayAt
BibTex
@InProceedings{AalMuh_SubspaceBased_MICCAISAT2026,
author = { Aaliyan, Muhammad AND Munir, Mustafa AND Bhattarai, Binod AND Gyawali, Prashnna Kumar},
title = { { Subspace-Based Auditing of Bottlenecks in Retinal Fundus Foundation Embeddings } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}