Agentic AI Framework Boosts Glaucoma Detection Accuracy

Large language models (LLMs) show potential in medical image interpretation but are often hampered by issues such as generating incorrect information (hallucination), limited accuracy, and inconsistent results across different runs. These limitations are particularly critical in high-stakes medical applications like disease detection. Researchers have developed and validated an agentic AI framework specifically for glaucoma detection from fundus photography, designed to mitigate these LLM shortcomings. The framework operates in three steps: an initial LLM assessment, followed by function calls to specialized deep learning tools for image quality assessment, glaucoma classification, and optic disc/cup segmentation, and finally, an LLM reflection phase that integrates the initial impression with the tool outputs. This agentic workflow dramatically improved classification accuracy by 16 to 47 percentage points, bringing it within 6 points of a fellowship-trained glaucoma specialist, and even matching specialist accuracy on one dataset. It corrected LLM-alone issues like positive bias and stochastic variability, significantly reducing cup-to-disc ratio error and boosting run-to-run consistency. This suggests a shift towards orchestrated multi-agent systems in medical AI, combining LLM reasoning with specialized, reliable tools.