Deep CNNs for Diabetic Retinopathy Classification: A Transfer Learning Perspective
Ruthran Baskar, Sabu Emmanuel, Claudia Mazo
Abstract
Diabetics Retinopathy is a very common eye disease among diabetic patients that affects the blood vessels in the retina, around 3.9 million people are estimated to be affected by diabetic retinopathy globally. Early detection and accurate diagnosis are crucial for timely intervention and management of the disease. On the other hand, in the past few years, deep learning has shown great success in the medical field, and transfer learning is one of the most potent techniques among them. To detect and classify the five diabetic retinopathy stages —normal, mild, moderate, severe, and Proliferative Diabetic Retinopathy (PDR)—, we have trained Alexnet and DenseNet-169 architectures using the APTOS2019 and the Diabetic Retinopathy Competition datasets. Both architectures were tuned on 20,163 images (9000 normal, 2808 mild, 6287 moderate, 1065 severe, and 1003 PDR images) and tested on 2017 images (900 normal, 281 mild, 629 moderate, 107 severe, and 100 PDR images). Among the two architectures, DenseNet-169 showed an overall better result in classifying each stage of diabetic retinopathy. DenseNet169 obtained an F1-score of 0.55, 0.38, 0.40, 0.60, and 0.69 for normal, mild, moderate, severe, and PDR, respectively. This study highlights the potential of transfer learning in improving diabetic retinopathy classification, contributing to early diagnosis and effective management of the disease, and ultimately enhancing patient care and outcomes.