Development of an ensemble transfer learning-based convolutional neural networks model for grading of diabetic retinopathy

Abstract

Diabetic Retinopathy (DR) is one of the diseases that infect people who suffer from diabetes. This chronic disease harms the patient retina and is considered one of the main causes of total blindness for people in the mid-age. Diagnosis of this disease is time-consuming and not accessible in some countries where the number of patients is very big comparing to the number of ophthalmologists. Therefore, designing and developing automated systems to grade DR is considered one of the recent research areas in the world of medical image applications. In this research, a complete pipeline for retinal fundus images processing and analysis was described, implemented and evaluated. This pipeline has three main stages: (i) image pre-processing, (ii) features extraction and (iii) classification. In the first stage, the image was pre-processed using different transformations techniques. In the second stage, the convolution neural network algorithm (CNN) was used. The concept of transfer learning and fine-tuning were advocated in this research. ResNet, DenseNet, and SqueezeNet were fine-tuned in order to implement the features extraction stage. For the classifier in the last stage, decision tree-based algorithms with the concept of ensemble learning were used where Random Forest, XGBoost and LightGBM were implemented and evaluated. Kaggle diabetic retinopathy dataset, a publicly available dataset, of retinal fundus image was used for training and testing. The problem of DR diagnosis was handled as a multi-class classification problem where there are five levels of the disease severity (0 – No DR, 1 – Mild, 2 – Moderate, 3 – Severe, 4 – Proliferative DR). The final model developed in this research used ResNet101 and DenseNet169 for features extraction, and it used the XGBoost for classification. It produced a very accurate performance with a quadratic weighted kappa score of 91.4% and an accuracy of 96.5%. This research proves that using CNN as a features-extractor algorithm is highly efficient since it produced representative features for the used images dataset. It shows that using the imbalanced dataset sampler is a very efficient solution to handle the issue of the imbalanced dataset. Also, it proves that ensemble learning algorithms are very promising algorithms to be used since they produced a very accurate model. The final model developed in this research could be used as the main unit for a computer-aided system (CAD) to be hosted online for DR diagnosis.