Machine learning and diabetic retinopathy: A comparative study of convolutional neural network models for diabetic retinopathy classification

Sammanfattning

Diabetic Retinopathy (DR) is a leading cause of preventable blindness on a global scale. Withan increasing prevalence of diabetes, manual screening methods are therefore becoming increasinglyresource-intensive. This study evaluates the efficacy of Convolutional Neural Networkswhen applied to automated classification of DR severity. The study used a combined datasetderived from combining APTOS 2019 and EyePacs. A comparative analysis was conductedon the performance of the three architectures: VGG16, ResNet50 and InceptionV3 by using atwo-stage experimental design. Phase I (Baseline) assessed the intrinsic feature extraction capabilitiesof each respective model, while Phase II (Enhanced) applied cost-optimizing learningalongside algorithmic regularization such as MixUp. The results present that standard architecturesstruggle to generalize and classify DR on imbalanced data, whereas VGG16 completelyfailed to identify several classes of the five point severity grade scale. The subpar performancedemonstrated that imbalanced medical data requires targeted optimization techniques. The performancemetrics across both phases are; VGG16 achieved a quadratic kappa value (QWK)of 𝜅 = 0.159, the optimized ResNet50 model achieved a QWK score of 0.650 and a recall of57.3% for Proliferative DR, respectively the baseline achieved 41.3% in the same category. Furthermore,InceptionV3 displayed similar performance to that of ResNet, but still slightly lowerover most of the categories evaluated. The study concludes that while resolution constraints(256×256, 299×299) pose as a bottleneck to achieve state-of-the-art performance, the proposedoptimization protocol provide a good starting point for future research, along with significantlyincreasing the model robustness and thus validating the potential application of CNNs as anassistant in clinical workflows.