همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Reliability-Aware Prompting for Diabetic Retinopathy Classification using Vision-Language Models

نویسندگان :

Danial Zohourian¹ Dorsa Asgari² Sadegh Madadi³ Hadi Farahani⁴

1- Shahid Beheshti University 2- Shahid Beheshti University 3- Amirkabir University of Technology 4- Shahid Beheshti University

کلمات کلیدی :

diabetic retinopathy،artificial intelligence،vision-language models،prompt ensemble

چکیده :

Diabetic Retinopathy (DR) remains one of the leading causes of preventable blindness, and automated grading of fundus images plays a crucial role in early diagnosis and treatment. Vision–language models (VLMs), such as CLIP and its medical variant FLAIR, have recently shown promise for zero-shot medical image classification without task-specific training. However, their performance is highly sensitive to prompt phrasing, and optimal label formulations vary across disease stages. In this study, we evaluate the FLAIR model on the APTOS dataset for both binary and five-class DR classification. After establishing a strong baseline, we introduce Reliability-Aware Prompting (RAP), a lightweight method that identifies prompt sets demonstrating high precision for specific DR stages on a training split and selectively applies them at inference time. Without modifying model parameters, RAP improves classification reliability by leveraging class-specific prompt trustworthiness. The results highlight the importance of prompt selection in medical VLMs and suggest a simple, training-free path to more dependable diagnostic performance.