Training and validation of an automated algorithm to differentiate no and minimal diabetic retinopathy from more severe stages in wide-field images.

PURPOSE: Diabetic retinopathy (DR) is a leading cause of blindness in the working-age population. Screening is essential to identify and treat sight-threatening stages prior to irreversible visual loss. This study aimed to train and validate an automated algorithm to identify no or minimal DR, potentially saving resources for specialist evaluation. METHODS: We included 307 wide-field images (Optomap, Dunfermline, Scotland) classified by a certified retinal expert according to the International Clinical Diabetic retinopathy (ICDR) scale. The expert manually annotated 26 995 diabetic-related retinal lesions using the Computer Vision Annotation Tool (CVAT). A segmentation model was trained to detect lesions using a 70/15/15 split for training, tuning, and validation. A classification model used the outputs of the segmentation model as input features for a decision tree classifier, categorizing patients into Group 0 (DR level 0-1) or Group 1 (DR level 2-4). Classifier hyperparameters and input feature selection were optimized based on binary classification performance using cross-validation. The classifier was evaluated on 48 validation images and further validated with 200 images graded by up to four independent certified graders. RESULTS: The area under the curve was 0.94 for the 48 validation images, with specificity, sensitivity, and kappa values of 0.89, 0.93, and 0.83, respectively. For the 200 expert-validated images, the values were 0.91, 0.98, 0.82, and 0.79, respectively. CONCLUSION: The combined method of segmentation followed by feature count analysis shows promising results for binary DR classification in Optomap wide-field images without requiring a large dataset for model development.