Diabetic Retinopathy Classification: Performance Evaluation of Pre-trained Lightweight CNN using Imbalance Dataset

Abstract

Diabetic Retinopathy (DR) is an eye complication that arises from long-term diabetes and damages the retinal blood vessels. Various clinical studies claim that Diabetic retinopathy infects about eighty percent of patients who suffer from diabetes type 1 for the last 15 years and a hundred percent of patients with this disease for 20 years. The human evaluation method is challenging but useful because it can detect diseases by the presence of lesions associated with Diabetic Retinopathy in most cases, but it is also time-consuming, erroneous, and requires a sophisticated medical setup. An efficient and automatic Diabetic Retinopathy identification method is still a challenging task. The feature extraction part is a very significant part and plays a vital role in the automatic Diabetic Retinopathy identification system. CNN has demonstrated its efficiency in medical image classification tasks as compared to other neural networks and traditional image processing methods. In this study, two lightweight CNN models: MobileNet and MobileNetV2 are used via transfer learning for binary (2-class) and multiclass (5-class) Diabetic Retinopathy classification using the DDR dataset, which is highly imbalanced. The efficiency of the models is measured using accuracy, precision, recall, and F1-score values. The ROC curve is generated for both models in binary and multiclass classification. The MobileNet model performed slightly better than MobilenetV2 in Diabetic Retinopathy classification for binary and multiclass classification. MobileNet shows 80% and 71% accuracy whereas MobileNetV2 shows 79% and 69% in binary and multiclass classification, respectively.