COUNTERFACTUAL LESION REASONING BASED EVIDENCE VERIFIED RAG FOR EXPLAINABLE DIABETIC RETINOPATHY SEVERITY CLASSIFICATION AND CLINICAL REPORT GENERATION

Authors

  • Dr. N. Balakumar Author
  • Vinitha k Author

DOI:

https://doi.org/10.4238/2p801050

Keywords:

Diabetic Retinopathy, Severity Classification, Counterfactual Reasoning, Retrieval-Augmented Generation, Evidence Verification, Clinical Report Generation

Abstract

Early diagnosis and robust severity classification are significant for providing proper treatment and preventing further damage to vision. Therefore, this research proposes a Counterfactual Lesion Reasoning based Evidence-Verified Retrieval-Augmented Generation (CLR-EVRAG) framework with aiming at enhancing performance of diabetic retinopathy severity classification and clinical report generation. First, retinal fundus images are preprocessed in CLR EVRAG to find out significant abnormalities, such as microaneurysms, retinal hemorrhages, and hard exudates. The Counterfactual Lesion-Consistency Ordinal Network (CLCON) is then utilized in CLR-EVRAG to classify DR into different severity levels. A counterfactual reasoning mechanism is utilized in CLR-EVRAG to analyze how changes in relevant lesion evidence affect the predicted severity. This process helps assess whether the model's decisions are meaningfully connected to the detected retinal abnormalities and identifies cases that require further investigation. Next, the CLR-EVRAG uses Retrieval-Augmented Generation to obtain relevant information from a carefully selected collection of trusted clinical guidelines and medical references. The retrieval process considers the detected lesions, predicted severity, and evidence required to explain the classification. The retrieved information is then combined with the image-based findings to generate a structured clinical report. An evidence-verification module verifies whether significant statements in the generated report are supported by the retinal image, the retrieved medical sources, or both. Unsupported or uncertain statements are flagged for expert review. The performance of CLR-EVRAG is determined using classification accuracy, retrieval precision, computational complexity, and false positive rate.

Downloads

Published

2026-11-04

Issue

Section

Articles