CauDR: A causality-inspired domain generalization framework for fundus-based diabetic retinopathy grading.

Area of Science:

• Ophthalmology and Artificial Intelligence
• Medical Imaging Analysis
• Computational Causality

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss, necessitating accurate and efficient screening.
• Computer-aided DR grading systems are crucial for rapid diagnosis, but deep learning models struggle with generalization across different imaging domains.
• Domain shifts caused by variations in imaging devices and protocols lead to performance degradation in current DR grading algorithms.

Purpose of the Study:

• To develop a novel deep learning framework for diabetic retinopathy grading that overcomes limitations in cross-domain generalization.
• To investigate the role of spurious correlations in DR grading model performance degradation.
• To introduce a causality-inspired approach for more robust and generalizable DR diagnostics.

Main Methods:

• Developed a universal structural causal model (SCM) to identify and analyze spurious correlations in fundus images.
• Proposed CauDR, a causality-inspired framework designed to eliminate spurious correlations.
• Reorganized existing datasets into the 4DR benchmark for domain generalization (DG) scenarios.

Main Results:

• The proposed CauDR framework demonstrated significant improvements in generalizability across different domains.
• CauDR achieved state-of-the-art (SOTA) performance in diabetic retinopathy grading tasks.
• The causality-inspired approach effectively mitigated performance decline caused by domain shifts.

Conclusions:

• Incorporating causal analysis into deep learning models can enhance the generalizability of DR grading systems.
• CauDR offers a promising solution for developing more reliable and robust AI-based diagnostic tools for diabetic retinopathy.
• The 4DR benchmark facilitates further research in domain generalization for medical imaging analysis.