Domain Generalization for Medical Image Analysis: A Review

TL;DR

The paper tackles the challenge of domain shift in medical image analysis by providing a system-wide review of domain generalization (DG) techniques tailored to MedIA. It introduces a four-level taxonomy—data-level, feature-level, model-level, and analysis-level—and maps these methods onto the full MedIA workflow, from data acquisition to analysis. The authors analyze the strengths and limitations of each approach, discuss extreme source/target constraints, and propose future directions including medical foundation models and standardized benchmarks. The work aims to guide researchers and engineers in building robust, transferable MedIA systems with practical clinical impact. Overall, it highlights the importance of integrating DG across all stages of MedIA to improve reliability, safety, and generalizability in diverse clinical environments.

Abstract

Medical image analysis (MedIA) has become an essential tool in medicine and healthcare, aiding in disease diagnosis, prognosis, and treatment planning, and recent successes in deep learning (DL) have made significant contributions to its advances. However, deploying DL models for MedIA in real-world situations remains challenging due to their failure to generalize across the distributional gap between training and testing samples - a problem known as domain shift. Researchers have dedicated their efforts to developing various DL methods to adapt and perform robustly on unknown and out-of-distribution (OOD) data distributions. This article comprehensively reviews domain generalization (DG) studies specifically tailored for MedIA. We provide a holistic view of how DG techniques interact within the broader MedIA system, going beyond methodologies to consider the operational implications on the entire MedIA workflow. Specifically, we categorize DG methods into data-level, feature-level, model-level, and analysis-level methods. We show how those methods can be used in various stages of the MedIA workflow with DL equipped from data acquisition to model prediction and analysis. Furthermore, we critically analyze the strengths and weaknesses of various methods, unveiling future research opportunities.

Figures (5)

• Figure 1: Overview of the medical imaging analysis (MedIA) pipeline illustrating various stages and their associated domain generalization (DG) techniques. Stages include data acquisition, image reconstruction, upstream feature extraction, downstream task, and analysis. Each stage is associated with references to specific sections ($\S$), pages (p.), or external citations where the techniques are detailed.
• Figure 2: Number of publications per year on the Google Scholar database.
• Figure 3: Settings of source and target domain for domain generalization. Source domain consists of $M$ domains, where $M=1$ refers to single-source and $M>1$ refers to multi-source settings. The target domain consists of $K$ domains. Cross-site, cross-sequence, and cross-modality define unique settings of the target domain for MedIA. Cross-site refers to generalization to different sites, e.g., different devices, and healthcare facilities. Cross-sequence refers to generalization to different sequences, i.e., different acquisition times or imaging protocols. Cross-modality refers to generalization to different modalities, e.g., CT to MRI.
• Figure 4: Hierarchical structure of the different aspects of domain generalization (DG) for medical image analysis (MedIA). This taxonomy divides the DG strategies into four primary levels: data-level, feature-level, model-level, and analysis-level, each encompassing distinct sub-strategies.
• Figure 5: Problem-specific suggestion for strategies for integrating domain generalization into MedIA workflow. Diamond box indicates the start terminator, angled boxes indicate the process, and round boxes indicate the decision.