[Paper Review] A Survey on Domain Generalization for Medical Image Analysis

Ziwei Niu, Shuyi Ouyang+3 more|arXiv (Cornell University)|Feb 7, 2024

AI in cancer detection4 citations

TL;DR

This survey provides a comprehensive review of domain generalization (DG) methods in medical image analysis, addressing domain shift across different sites, modalities, and sequences. It categorizes existing approaches by data, feature, and model levels, details key algorithms, benchmarks datasets, and identifies future research directions, supported by a public GitHub repository with resources.

ABSTRACT

Medical Image Analysis (MedIA) has emerged as a crucial tool in computer-aided diagnosis systems, particularly with the advancement of deep learning (DL) in recent years. However, well-trained deep models often experience significant performance degradation when deployed in different medical sites, modalities, and sequences, known as a domain shift issue. In light of this, Domain Generalization (DG) for MedIA aims to address the domain shift challenge by generalizing effectively and performing robustly across unknown data distributions. This paper presents the a comprehensive review of substantial developments in this area. First, we provide a formal definition of domain shift and domain generalization in medical field, and discuss several related settings. Subsequently, we summarize the recent methods from three viewpoints: data manipulation level, feature representation level, and model training level, and present some algorithms in detail for each viewpoints. Furthermore, we introduce the commonly used datasets. Finally, we summarize existing literature and present some potential research topics for the future. For this survey, we also created a GitHub project by collecting the supporting resources, at the link: https://github.com/Ziwei-Niu/DG_for_MedIA

Motivation & Objective

• Address the challenge of domain shift in medical image analysis, where deep learning models degrade when deployed across different clinical sites, imaging modalities, or scan sequences.
• Systematically categorize and analyze recent domain generalization techniques in medical imaging into three levels: data manipulation, feature representation, and model training.
• Provide a detailed overview of representative algorithms at each level to guide researchers in selecting and developing robust DG methods.
• Compile and summarize commonly used medical imaging datasets for domain generalization research.
• Identify open challenges and propose promising future research directions for advancing generalization in clinical AI systems.

Proposed method

• Formally define domain shift and domain generalization in the context of medical image analysis, distinguishing between domain-invariant and domain-specific learning settings.
• Categorize DG methods into three levels: data-level (e.g., data augmentation, domain adaptation), feature-level (e.g., domain-invariant representation learning), and model-level (e.g., meta-learning, self-supervised training).
• Present detailed descriptions of key algorithms under each category, such as domain adversarial training, distribution alignment, and meta-optimization strategies.
• Review benchmark datasets used in DG for medical imaging, including their domains, modalities, and clinical applications.
• Analyze the strengths and limitations of each method category through comparative analysis of performance and generalization capacity.
• Host a public GitHub repository with curated resources, code links, and datasets to support reproducibility and further research.

Experimental results

Research questions

• RQ1 How do domain shift issues manifest in medical image analysis across different clinical settings and imaging protocols?
• RQ2 What are the key methodological categories for domain generalization in medical imaging, and how do they differ in design and performance?
• RQ3 Which algorithms and techniques have shown the most promise in improving model generalization across unknown domains in medical imaging?
• RQ4 What are the standard datasets used to evaluate domain generalization in medical image analysis, and how do they reflect real-world clinical variability?
• RQ5 What are the most pressing open challenges and future research directions in domain generalization for medical image analysis?

Key findings

• Domain generalization is essential for deploying deep learning models in real-world clinical settings where data distributions vary significantly across institutions and scanning protocols.
• Data-level methods such as domain augmentation and style transfer show effectiveness in improving robustness by simulating domain variations during training.
• Feature-level approaches, including domain-invariant representation learning and distribution alignment, achieve strong performance by minimizing domain-specific biases in learned features.
• Model-level strategies such as meta-learning and self-supervised pretraining demonstrate improved zero-shot generalization to unseen domains.
• The survey identifies a lack of standardized benchmarks and consistent evaluation protocols across studies, highlighting a need for unified evaluation frameworks.
• The authors' public GitHub repository aggregates key resources, including datasets, code, and models, to accelerate research and reproducibility in the field.