Med-SegLens: Latent-Level Model Diffing for Interpretable Medical Image Segmentation

Salma J. Ahmed, Emad Mohammed, Azam Asilian Bidgoli
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:1224-1241, 2026.

Abstract

Modern segmentation models achieve strong predictive performance but remain largely opaque, limiting our ability to diagnose failures, understand dataset shift, or intervene in a principled manner. We introduce $\textbf{Med-SegLens}$, a model-diffing framework that decomposes segmentation model activations into interpretable latent features using sparse autoencoders trained on SegFormer and U-Net. Through cross-architecture and cross-dataset latent alignment across healthy, adult, pediatric, and sub-Saharan African glioma cohorts, we identify a stable backbone of shared representations, while dataset shift is driven by differential reliance on population-specific latents. We show that these latents act as causal bottlenecks for segmentation failures, and that targeted latent-level interventions can correct errors and improve cross-dataset adaption without retraining, recovering performance in 70% of failure cases and improving Dice score from 39.4% to 74.2%. Our results demonstrate that latent-level model diffing provides a practical and mechanistic tool for diagnosing failures and mitigating dataset shift in segmentation models.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-ahmed26a, title = {Med-{S}eg{L}ens: Latent-Level Model Diffing for Interpretable Medical Image Segmentation}, author = {Ahmed, Salma J. and Mohammed, Emad and Bidgoli, Azam Asilian}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {1224--1241}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/ahmed26a/ahmed26a.pdf}, url = {https://proceedings.mlr.press/v306/ahmed26a.html}, abstract = {Modern segmentation models achieve strong predictive performance but remain largely opaque, limiting our ability to diagnose failures, understand dataset shift, or intervene in a principled manner. We introduce $\textbf{Med-SegLens}$, a model-diffing framework that decomposes segmentation model activations into interpretable latent features using sparse autoencoders trained on SegFormer and U-Net. Through cross-architecture and cross-dataset latent alignment across healthy, adult, pediatric, and sub-Saharan African glioma cohorts, we identify a stable backbone of shared representations, while dataset shift is driven by differential reliance on population-specific latents. We show that these latents act as causal bottlenecks for segmentation failures, and that targeted latent-level interventions can correct errors and improve cross-dataset adaption without retraining, recovering performance in 70% of failure cases and improving Dice score from 39.4% to 74.2%. Our results demonstrate that latent-level model diffing provides a practical and mechanistic tool for diagnosing failures and mitigating dataset shift in segmentation models.} }
Endnote
%0 Conference Paper %T Med-SegLens: Latent-Level Model Diffing for Interpretable Medical Image Segmentation %A Salma J. Ahmed %A Emad Mohammed %A Azam Asilian Bidgoli %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-ahmed26a %I PMLR %P 1224--1241 %U https://proceedings.mlr.press/v306/ahmed26a.html %V 306 %X Modern segmentation models achieve strong predictive performance but remain largely opaque, limiting our ability to diagnose failures, understand dataset shift, or intervene in a principled manner. We introduce $\textbf{Med-SegLens}$, a model-diffing framework that decomposes segmentation model activations into interpretable latent features using sparse autoencoders trained on SegFormer and U-Net. Through cross-architecture and cross-dataset latent alignment across healthy, adult, pediatric, and sub-Saharan African glioma cohorts, we identify a stable backbone of shared representations, while dataset shift is driven by differential reliance on population-specific latents. We show that these latents act as causal bottlenecks for segmentation failures, and that targeted latent-level interventions can correct errors and improve cross-dataset adaption without retraining, recovering performance in 70% of failure cases and improving Dice score from 39.4% to 74.2%. Our results demonstrate that latent-level model diffing provides a practical and mechanistic tool for diagnosing failures and mitigating dataset shift in segmentation models.
APA
Ahmed, S.J., Mohammed, E. & Bidgoli, A.A.. (2026). Med-SegLens: Latent-Level Model Diffing for Interpretable Medical Image Segmentation. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:1224-1241 Available from https://proceedings.mlr.press/v306/ahmed26a.html.

Related Material