Abstract
Breast MRI segmentation methods frequently fail to generalize across scanner manufacturers, with domain-induced performance degradation posing a critical barrier to reliable clinical deployment. We propose an out-of-distribution (OOD)-guided continual learning framework that enables autonomous adaptation to evolving scanner distributions without scanner metadata or manual domain labels. The method combines feature-space Mahalanobis-based OOD detection with unsupervised pseudo-domain discovery and rehearsal-based continual adaptation. Incoming MRIs are selectively flagged as OOD using a frozen source-trained encoder, clustered into emerging pseudo-domains, and used to trigger targeted fine-tuning with a multi-criteria exemplar memory buffer. We evaluate the framework on the MAMA-MIA dataset across two sequential adaptation cycles spanning three MRI scanner manufacturers. The model achieves an overall DSC of 0.697, with minimal forgetting (backward transfer = −0.014) and positive transfer to unseen domains (forward transfer = +0.047). Critically, OOD guidance reduces annotation burden by forwarding only 31.2% of incoming examinations for adaptation, while recovering four of five lesions missed by the baseline model on images acquired by a previously unseen scanner manufacturer. These results demonstrate that selective OOD-guided continual learning enables both efficient and clinically meaningful generalization under heterogeneous and evolving MRI acquisition conditions.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/Deep_Brea3th_028.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=NQVPi8QAa3
BibTex
@InProceedings{ForYas_OODGuided_MICCAISAT2026,
author = { Forghani, Yasna AND Andrade, Matilde AND Timóteo, Rafaela AND Loução, Nuno AND Marques, Tiago AND Figueiredo, Mário AND Gouveia, Pedro AND Santinha, João},
title = { { OOD-Guided Continual Learning for Multi-Scanner Breast MRI Lesion Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17256},
month = {pending},
page = {pending}
}