Diabetic Retinopathy (DR) is a major cause of vision loss and blindness in diabetic individuals. DR is conventionally diagnosed by assessing retinal lesion findings from fundus photographs taken during exams and applying a scale like International Classification of Diabetic Retinopathy (ICDR). The expected rise in future DR cases highlights the need for deep learning models capable of identifying relevant lesions and delivering explainable results. To this end we present BigEye, a novel framework that uses extracted lesion features to predict ICDR stage. A dataset of fundus images from a local hospital and a public dataset, annotated with segmentation masks and DR stages, is assembled to train a DeepLabV3 + model on six retinal lesions. Lesion quantities and pixel area features are integrated by a classifier model evaluated through 10-fold nested cross validation (0.77 ± 0.07 precision, 0.71 ± 0.06 recall, 0.72 ± 0.07 F1 score, 0.95 ± 0.02 ROC-AUC, 0.83 ± 0.03 accuracy). A Shapely Additive Explanations (SHAP) value analysis notably shows close alignment between discriminative lesions for each DR stage and corresponding ICDR stage criteria. These results demonstrate that BigEye is well suited for providing explainable ICDR stage predictions grounded in clinical knowledge.
Gill et al. (Sun,) studied this question.