Geração Automática de Laudos de Glaucoma por LLMs: Estudo Experimental Baseado em Exames de Campimetria computadorizada

Abstract

According to the World Health Organization (WHO), glaucoma is one of the leading causes of irreversible blindness worldwide, making early diagnosis essential to reduce disease progression. In this context, the automatic generation of medical reports using Large Language Models (LLMs) has the potential to support specialists and optimize clinical workflows. This study proposes the use of LLMs for the automatic generation of glaucoma reports based on computerized visual field examinations. Three language models—Llama-3.3-70B-Versatile, GPT-OSS-20B, and Gemini-2.5-Flash—were evaluated using a dataset of 118 examinations represented in JSON format. The generated reports were compared with reference reports using the BLEU, ROUGE-1, ROUGE-L, Semantic Similarity metrics, and the LLM-as-a-Judge approach. The results showed that the Llama-3.3-70B-Versatile model achieved the best overall performance, presenting the lowest average report generation time and the highest semantic similarity to the reference reports. Furthermore, the combination of automatic evaluation metrics and qualitative assessment provided a comprehensive analysis of the generated reports, demonstrating the potential of LLMs as a support tool for ophthalmological report generation.