Automatic classification of retinal diseases with transfer learning-based lightweight convolutional neural network

Abstract

Diabetic retinopathy (DR) and diabetic macular edema (DME) are the major causes of permanent blindness in the working-age population. Deep learning methods have been proposed to automatically grade DR and DME for ophthalmologists' design of tailored treatments for patients. However, these methods are computationally intensive with a large number of parameters and affect the optimization of hyperparameters, making them challenging to deploy to mobile or embedded devices with limited computer resources. In this paper, we developed a transfer learning-based lightweight convolutional neural network to jointly classify the severity of DR and DME. Using fivefold cross-validation, our model achieved an average accuracy, precision, recall, specificity, and F1-score of 0.9666, 0.9700, 0.9685, 0.9932, and 0.9663, respectively, better than MobileNet V2, while the number of parameters and the recognition speed were dramatically less than those of MobileNet V2 and ResNet50. These results show that our model is hopeful in diagnosing retinopathy in clinical trials, even when configured for mobile and embedded devices.