EYENOVA: AN EXPLAINABLE DEEP LEARNING PROTOTYPE FOR DIABETIC RETINOPATHY SEVERITY GRADING AND LESION LOCALIZATION
Abstract
Diabetic retinopathy is a major cause of preventable vision loss, and timely screening is essential because early disease stages may remain asymptomatic. This study presents EyeNova, a locally deployable web-based prototype for diabetic retinopathy screening support using retinal fundus images. The prototype combines image validation, preprocessing, image-quality assessment, five-class DR severity grading, exploratory lesion localization, Grad-CAM visualization, dashboard presentation, and structured report generation. For severity grading, an EfficientNetB0 model initialized with ImageNet weights was trained on the Kaggle Diabetic Retinopathy 224x224 Gaussian Filtered dataset using a class-folder-wise 80/20 random train–validation split with seed 42, resulting in 2,805 training images and 704 validation images across five DR severity classes. For lesion localisation, a YOLOv8n detector was trained on a YOLO-format reconstruction of the FGADR Seg-set pixel-level lesion masks, using an image-level 80/20 stratified train-validation split by DR grade with 1,476 training images, 366 validation images, and 103,354 lesion annotations across six lesion classes. On the available internal validation split, the classifier achieved 98.44% accuracy, 98.45% weighted precision, 98.44% weighted recall, 99.46% weighted specificity, and 98.43% weighted F1-score, with an average inference time of 24.61 ms per image. In contrast, lesion localisation remained the limiting subsystem, with 59.70% precision, 21.27% recall, 18.44% mAP@0.50, and 7.30% mAP@0.50:0.95. These findings indicate that EyeNova is promising as a preliminary explainable screening-support prototype for severity grading, but its evidence remains limited to internal validation, with no independent external test set currently available. Accordingly, EyeNova should not be interpreted as a clinically validated diagnostic system.
Keywords
- Diabetic retinopathy, fundus imaging, EfficientNetB0, YOLOv8n, lesion localisation, explainable artificial intelligence, Grad-CAM, screening-support system.
Views & Downloads Statistics (Last 12 Months)
Monthly performance and trends of article views and downloads
