Abstract

Scaling deep learning models across evolving clinical environments requires robust Continual Learning (CL) mechanisms to overcome severe domain shifts. However, standard visual CL approaches in medical imaging frequently suffer from domain overfitting and shortcut learning, where models memorize spurious physical artifacts rather than genuine biological pathology. Furthermore, visual-only dynamic routing often exacerbates this by causing negative transfer in out-of-distribution clinical scenarios. To address the stability-plasticity dilemma under severe clinical covariate shift, we propose a Multimodal-Guided Continual Adaptation framework. Our approach decouples 3D visual representations from tabular clinical metadata using a dual-stream architecture. To enable Domain-Incremental Learning (Domain-IL), we freeze a generalized visual backbone and sequentially train domain-specific Deep Latent Adapters. Crucially, the framework leverages the tabular modality to compute a demographic sparsity metric. This multimodal gating dynamically routes common patient profiles to specialized visual adapters, while deploying a cautious cross-domain ensemble to prevent negative transfer for atypical clinical profiles. Additionally, an entropy-weighted prior correction mechanism utilizes the tabular stream to mathematically subtract hardware-specific priors directly from the visual predictions. Evaluations on a multi-hospital 3D CT continual learning benchmark demonstrate that our multimodal fusion strategy effectively preserves representation stability, ensures target plasticity, and neutralizes domain-specific biases across sequential clinical deployments.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CLiMeM_018.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=xeA25gEbO5&referrer=%5BProgram%20Chair%20Console%5D(%2Fgroup%3Fid%3DMICCAI.org%2F2026%2FWorkshop%2FCLiMeM%2FProgram_Chairs%23submission-status)

BibTex

@InProceedings{ThePar_Multimodal_MICCAISAT2026,
        author = { Theofilou, Paraskevi-Antonia AND Tzouveli, Paraskevi AND Kollias, Stefanos},
        title = { { Multimodal Fusion for Domain-Incremental Continual Learning in 3D Medical Imaging } },
        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}
}


back to top