Hybrid Convolutional Transformer Learning Utilizing Ordinal Sensitive Loss for Automated Grading of Diabetic Retinopathy

Abstract

Diabetic retinopathy (DR) is a major cause of preventable blindness globally requiring precise and dependable automated grading systems to facilitate extensive screening initiatives. Recent deep learning techniques utilizing convolutional neural networks (CNNs) have yielded encouraging outcomes however, they predominantly concentrate on local lesion identification and frequently neglect to encompass the global retinal context. Furthermore, the majority of current methodologies regard diabetic retinopathy grading as a conventional multi-class classification issue, neglecting the ordinal characteristics of disease severity and the significant class imbalance present in clinical datasets. In this paper, we propose a hybrid convolutional-transformer learning framework with ordinal-sensitive loss for automated diabetic retinopathy grading. The suggested model combines a deep CNN backbone for strong local feature extraction with a lightweight transformer encoder that works on convolutionally down sampled feature maps to model long-range dependencies in a computationally efficient manner. This design reduces the quadratic complexity of self-attention while keeping the global context information needed to figure out how bad something is. To deal with class imbalance and penalties for misclassifying ordinal data, an ordinal-sensitive focal loss is used to make the model focus on clinically important mistakes. We test the framework on publicly available fundus image datasets using a wide range of performance metrics, such as accuracy, macro-F1 score, balanced accuracy, area under the ROC curve (AUC), and quadratic Cohen's kappa. Experimental results show that the suggested method consistently beats CNN-only baselines and standard cross-entropy-based training, getting better accuracy of 85%