Agentic AI Framework Enhances Glaucoma Detection from Fundus Photography

A novel framework for agentic AI has been created, merging large language models (LLMs) with targeted deep learning technologies to enhance the detection of glaucoma through fundus photography. This framework tackles key issues associated with LLMs, such as hallucinations, inconsistent accuracy, and variability between runs. The process unfolds in three phases: an LLM conducts an initial evaluation, specialized tools are activated for assessing image quality (QAModel, FundaQ-8), classifying glaucoma (SwinV2-Tiny), and segmenting the optic disc/cup (SegFormer-B0), followed by a reflection phase where the LLM combines its initial assessment with the outputs from the tools. Evaluation occurred on two public datasets, ORIGA (n=100) and RIM-ONE-v3 (n=100), using both cropped and uncropped views. Testing involved two LLMs, Gemini 2.5 Flash and GPT-5.4 mini, with all images rated by a masked glaucoma specialist. The agentic framework enhanced classification accuracy by 16 to 47 percentage points across both datasets. This study, identified as arXiv:2608.07651, illustrates a promising pathway for improving LLM reliability in interpreting medical images, potentially resulting in more precise and consistent diagnostic tools in ophthalmology.