Single Domain Generalization for Rare Event Detection in Medical Imaging

Payal Kamboj, Ayan Banerjee, BIN XU, Sandeep Gupta

Keywords: Deep Learning, Knowledge, Rare Event Detection, Out-of-distribution detection

Abstract: Single Domain Generalization (SDG) addresses the challenge of training a model on a single domain to ensure it generalizes well to unseen target domains. Although extensively studied in image classification, there is a lack of prior work on SDG for rare event or image classification in imbalanced dataset. In the medical diagnosis and disease detection domain, where data is often limited and events of interest are rare, deep learning (DL) models frequently exhibit suboptimal performance, leading to poor generalization across datasets. In multi-center studies, disparate data sources, differences in scanners and imaging protocols introduce domain shifts that exacerbate variability in rare event characteristics. This paper addresses this challenge by first leveraging a pre-trained large vision model to rank classes based on their similarity to the rare event class, allowing focused handling of the most similar class, and then integrates domain-invariant knowledge on rare event with DL to accurately classify the rare event class. By carefully incorporating expert knowledge with data-driven DL, our technique effectively regularizes the model, enhancing robustness and performance even with limited data availability. We present a case study on seizure onset zone detection using fMRI data, demonstrating that our approach significantly outperforms state-of-the-art vision transformers, large vision models, and knowledge-based systems, achieving an average F1 score of 90.2% while maintaining an overall F1 score of 85.0% across multi-center datasets.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11841