A fine-grained fundus image dataset for cataract severity assessment and diagnosis

Cataracts are a leading global cause of blindness. While deep learning has shown promise in aiding cataract diagnosis, existing methods typically provide only binary assessments. Clinical practice, however, requires multi-level severity assessment and corresponding diagnostic recommendations. A significant limitation has been the lack of high-quality, publicly available datasets with fine-grained cataract classifications and associated professional diagnostic descriptions. To address this, we constructed the Cataract Severity and Diagnostic Image Dataset (CSDI), comprising 187 fundus image cases with detailed clinical diagnostic reports. We also propose a Multimodal Large Language Model (MLLM)-based diagnostic framework capable of fine-grained severity scoring from fundus images and generating professional diagnostic text. This study, to our knowledge, is the first to apply MLLM technology for precise quantitative cataract diagnosis. We benchmarked several state-of-the-art MLLMs on CSDI for severity assessment and diagnosis report generation, which confirmed high quality of our annotations. The CSDI dataset and associated annotations are publicly available.