Abstract
<title>Abstract</title> <p>Diabetic retinopathy (DR) is a common cause of vision loss among working-age adults, and DR treatment planning requires multidisciplinary collaboration across ophthalmology and endocrinology. In routine practice, limited specialist resources, documentation burden, and the difficulty of implementing multidisciplinary team consultation prevent many patients from receiving individualized treatment plans. We present three linked components: a large-scale, publicly available DR treatment-planning dataset, linking fundus images, clinical records, and treatment plans; DRAgent, a role-based multi-agent system in which clinical agents collaborate to generate individualized plans; and an evaluation framework combining guideline compliance, objective metrics, and human evaluation. DRAgent outperforms the corresponding standalone LLMs on nearly all metrics, raising human evaluation scores from 4.02 to 4.80/5. Ablation studies show that removing any specialist agent degrades performance in its clinical domain, confirming complementary roles. By turning multidisciplinary decision-making into an evaluable system, DRAgent may provide a useful starting point for developing similar systems in other diseases requiring cross-specialty care.</p>