AI models offer interpretable diabetic retinopathy grading with visual and text explanations

Researchers have developed a new method for grading diabetic retinopathy (DR) that combines deep learning models with interpretable explanations. The approach uses CNN and transformer architectures, achieving a QWK score of up to 0.934 through weighted soft voting ensembling. For interpretability, the study generated visual attribution maps using Grad-CAM++ and textual rationales from vision-language models, aiming to provide clinically meaningful insights from retinal images. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Enhances interpretability of medical AI models, potentially improving clinical trust and adoption for DR grading.

RANK_REASON This is a research paper detailing a new methodology for medical image analysis and interpretability.

Read on arXiv cs.CV →

• Diabetic Retinopathy
• CNN
• Transformer
• Grad-CAM++
• Vision-Language Models
• ResNet-50
• ConvNeXt-Tiny
• APTOS 2019
• arXiv

• paper
• model release