A novel few-shot classification framework for diabetic retinopathy detection and grading

Diabetes Retinopathy (DR) is a major microvascular complication of diabetes. Computer-Aided Diagnosis (CAD)
tools for DR management are primarily developed using Artificial Intelligence (AI) methods, such as machine and
deep learning algorithms. DR diagnostic tools have been developed in recent years using deep learning models. Thus, these models require large amounts of data for training. Consequently, these huge amounts of data are not balanced due to fewer cases in the dataset. To solve the problems associated with training models with small datasets, such as overfitting and poor approximation, this paper proposes a paradigm called Few-Shot Learning (FSL) which uses a relatively small amount of training data to train the models effectively. This paper proposes a novel prototype network, a type of FSL classification network capable of grading and detecting DR based on attention. The DRNet framework uses episodic learning to train its model on few-shot classification tasks. We developed a DRNet based on the APTOS2019 dataset for diabetic detection and grading. In the proposed network, aggregated transformations and gradient activations of classes are leveraged to design the attention
mechanism to capture image representations. As a result, the system achieves 99.73 % accuracy, 99.82 % sensitivity, 99.63 % specificity in DR detection, 98.18 % accuracy, 97.41% sensitivity, and 99.55% specificity in DR grading. An analysis of objective performance metrics and model interpretation shows that the proposed model can detect DR more efficiently and grade the severity more accurately when using unseen fundus images than existing state-of-the-art methods. Therefore, this tool could help provide a second opinion to an ophthalmologist about the severity level of DR.