Knowledge distillation-based lightweight MobileNet model for diabetic retinopathy classification

Diabetic retinopathy (DR) stands as a leading cause of global blindness. Early identification and prompt treatment are essential to prevent vision impairment caused by DR. Manual screening of retinal fundus images is challenging and time-consuming. Additionally, in low-income countries, there is a significant gap between the number of DR patients and ophthalmologists. Currently, machine learning (ML) and deep learning (DL) are becoming a viable alternative to traditional DR screening techniques. However, DL suffers a major limitation in resource-constrained devices because of its large model size and substantial computational demands. Knowledge distillation is a prominent technique for creating lightweight models, effectively transferring knowledge from a larger, complex model to a smaller, more efficient one without significant loss in performance. Therefore, in this research, a lightweight student model is proposed, which follows the MobileNet architectural design by utilizing depthwise separable convolutions. This design ensures efficient performance suitable for edge device deployment. For binary classification, our proposed model achieved an accuracy, precision, and recall of 98.38% on the APTOS 2019 dataset, whereas the proposed model achieved an accuracy of 93.03% for ternary classification on APTOS 2019.