Abstract
Continual learning (CL) in medical imaging must address sequential hospital-level domain shifts under strict memory constraints. We propose an active rehearsal framework that optimizes memory buffer population through two complementary strategies: (1) an uncertaintyaware method using Monte Carlo Dropout to disentangle aleatoric and epistemic uncertainty into a robust informativeness score, and (2) an unsupervised dissimilarity strategy maximizing latent feature diversity without requiring selection labels. We evaluate our approach on three domain-incremental scenarios spanning distinct source (Task A) and target (Task B) domains: OrganMNIST, Camelyon17 (multi-center histopathology), and HITS-5K (a private multi-center transcranial Doppler ultrasound cohort). Results demonstrate that our targeted selection mechanisms achieve comparable or superior performance than regularization (e.g., EWC) and standard uniform replay baselines. Notably, on HITS5K, all replay strategies exhibit strong positive backward transfer, improving Task B accuracy beyond 93% while reducing forgetting of Task A to −17%. At higher memory ratios, performance converges, suggesting
that optimized minimal replay suffices for synergistic tasks. Moreover, qualitative analysis of memory coverage in the latent space validates
the theoretical motivations behind our sampling mechanisms. Our work demonstrates that adaptive, semantics-aware replay CL improves model
robustness for medical image analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CLiMeM_008.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{DupMat_Integrating_MICCAISAT2026,
author = { Dupouy, Mathilde AND Vindas Yassine, Yamil AND Bahirah, Novia Putri AND Dambry, Thibaut AND Guépié, Blaise Kévin AND Teodoro, Douglas AND Delachartre, Philippe},
title = { { Integrating Uncertainty and Latent Feature Diversity for Robust Memory Replay in Continual Medical Imaging Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17263},
month = {pending},
page = {pending}
}