RetiGen: Framework leveraging domain generalization and test-time adaptation for multi-view retinal diagnostics

Area of Science:

• Ophthalmic imaging
• Medical machine learning
• Computer vision

Background:

• Domain generalization (DG) and test-time adaptation (TTA) improve medical imaging model accuracy across different domains.
• Existing methods often address DG or TTA separately, not leveraging their combined strengths.
• Domain shift remains a significant challenge, hindering the performance of AI in diagnostics.

Purpose of the Study:

• To introduce RetiGen, a novel test-time optimization framework that integrates DG and TTA in an end-to-end fashion.
• To enhance the robustness and accuracy of machine learning models in the ophthalmic imaging domain, specifically using multi-view color fundus photographs.
• To address the persistent issue of domain shift in medical imaging analysis.

Main Methods:

• RetiGen is a test-time optimization framework designed for integration with existing domain generalization approaches.
• It utilizes unlabeled multi-view color fundus photographs, leveraging information from multiple viewing angles.
• The framework integrates class balancing, test-time adaptation, and a multi-view optimization strategy to combat domain shift.

Main Results:

• RetiGen significantly enhances the generalizability and accuracy of ophthalmic imaging models.
• On the MFIDDR dataset, RetiGen improved AUC from 0.751 to 0.872 (0.121 improvement).
• On the DRTiD dataset, RetiGen improved AUC from 0.794 to 0.879 (0.085 improvement), outperforming state-of-the-art methods.

Conclusions:

• RetiGen effectively addresses domain shift by combining DG and TTA in an end-to-end framework.
• The proposed method demonstrates superior performance compared to existing state-of-the-art techniques in both DG and TTA.
• RetiGen offers a promising solution for improving diagnostic accuracy in ophthalmic imaging and other medical domains.