A test-time clinically adaptive framework for detecting multiple fundus diseases harnessing ophthalmic foundation...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Fundus diseases cause significant vision impairment globally.
• Ophthalmic foundation models (FMs) show potential but face challenges in multi-disease detection.
• Clinical translation is hindered by data imbalance, prediction uncertainty, and domain shifts.

Purpose of the Study:

• To introduce RetExpert, a test-time clinically adaptive framework to enhance FMs for multi-fundus disease detection.
• To improve robustness, generalizability, and clinical viability of AI models for fundus disease screening.

Main Methods:

• RetExpert utilizes adaptive knowledge units and a stochastic one-hot activation module.
• Employs long-tail-aware and uncertainty-aware multi-label learning strategies.
• Integrates a fundus disease co-occurrence matrix and test-time adaptation (TTUL+TTPL).

Main Results:

• RetExpert demonstrates superior detection performance and reliability compared to existing ophthalmic FMs.
• The framework shows enhanced cross-domain adaptability across 15 datasets.
• Achieved dynamic parameter adjustment without full model retraining.

Conclusions:

• RetExpert offers a clinically viable solution for automated multi-disease screening from fundus photographs.
• The framework addresses key limitations of current AI models in ophthalmology.
• Provides a robust and generalizable approach for real-world vision impairment detection.