Learning Ophthalmologist Clinical Reasoning for Glaucoma Diagnosis from Fundus Images
TL;DR
A reasoning-driven vision-language framework that explicitly models the ophthalmologist's diagnostic workflow by generating structured clinical reasoning prior to diagnosis is developed, demonstrating that explicitly modeling expert clinical reasoning simultaneously improves interpretability and diagnostic performance.
Abstract
Glaucoma is a leading cause of irreversible blindness worldwide. Ophthalmologists diagnose glaucoma through a structured reasoning process by sequentially evaluating optic nerve head characteristics before reaching a final diagnosis, whereas existing AI systems typically perform direct image classification without providing clinically meaningful reasoning. We present the first clinically annotated fundus reasoning dataset, comprising 1,077 fundus photographs paired with expert-authored six-step diagnostic reports. Building on this dataset, we develop a reasoning-driven vision-language framework that explicitly models the ophthalmologist's diagnostic workflow by generating structured clinical reasoning prior to diagnosis. The generated reports are clinically validated, achieving the best performance across all evaluated clinical findings, including a cup-to-disc ratio mean absolute error of 0.070, an ISNT Kendall distance of 1.73, and the highest semantic agreement with expert reports (BERTScore-F1 = 0.874). The resulting framework also improves glaucoma diagnosis, achieving a balanced accuracy of $94.7%$ and precision of $94.8%$, demonstrating that explicitly modeling expert clinical reasoning simultaneously improves interpretability and diagnostic performance. Code and data are available at url{https://glaucoma-cot.github.io/}.