Explainable Deep Learning for Lesion-Level Detection of Diabetic Retinopathy: A Segmentation Approach Using Fundus Images Graded as Mild-to-Moderate Nonproliferative Diabetic Retinopathy
Authors
Affiliations (1)
Affiliations (1)
- Saitama Medical Center: Saitama Ika Daigaku Sogo Iryo Center
Abstract
Deep learning has shown promise in diabetic retinopathy screening using fundus images. However, many existing models operate as "black boxes," providing limited interpretability at the lesion level. This study aimed to develop an explainable deep learning model capable of detecting four diabetic retinopathy-related lesions--hemorrhages, hard exudates, cotton wool spots, and microaneurysms--and evaluate its performance using both conventional per-lesion metrics and a novel syntactic agreement framework. A total of 1,087 fundus images were obtained from publicly available datasets (EyePACS and APTOS), which contained 585 images graded as mild-to-moderate nonproliferative diabetic retinopathy (DR1 or DR2). All images were manually annotated for the presence of the four lesions. A U-Net-based segmentation model was trained to generate binary predictions for each lesion type. The performance of the model was evaluated using sensitivity, specificity, precision, and F1 score, along with five syntactic agreement criteria that evaluated the lesion-set consistency between the predicted and ground truth outputs at the image level. The model achieved high sensitivity and F1 scores for hemorrhages and hard exudates, showed moderate performance for cotton wool spots, and failed to detect any microaneurysms (0% sensitivity), with 92.9% of the microaneurysms cases misclassified as hemorrhages. Despite this limitation, the image-level agreement remained high, with any-lesion match and hemorrhage match rates exceeding 95%. These findings suggest that although individual lesion classification was imperfect, the model effectively recognized abnormal images, highlighting its potential as a screening tool. The proposed syntactic agreement framework offers a complementary evaluation strategy that aligns more closely with clinical interpretation and may help bridge the gap between artificial intelligence-based predictions and real-world ophthalmic decision-making.