Attention augmented feature fusion in a hybrid CNN–transformer for fine-grained diabetic retinopathy severity grading

Despite the advent of automated diabetic retinopathy (DR) severity grading from retinal fundus images, it remains challenging because of class imbalance, subtle DR lesion characteristics, and limited generalization of currently available deep learning models. Given the above, this study aims to fill these gaps by introducing a novel hybrid CNN-Transformer architecture (EffTNet), which extracts local lesion features by using the EfficientNet-B3a network and learns the global contextual features by using the ViT-B16 network. A novel Attention-Augmented Feature Fusion (AAFF) module is presented to adaptively combine complementary features from both branches by channel-wise attention and feature recalibration. Furthermore, pair-aware contrastive learning is used to boost inter-class separability and improve the recognition of the underrepresented severe DR categories. The harmonised dataset used for training the model included APTOS 2019, EyePACS and Messidor-2 images, while the other three datasets (DDR, DiaRetDB1 and IDRiD) were external and used for validation of the model. The experimental results show that the EffTNet obtained an accuracy of 98.77%, a precision of 96.30%, a recall of 97.83%, an F1-score of 96.99%, and a Quadratic Weighted Kappa score of 0.947. A grad-CAM analysis also shows that the model is trained on clinically relevant retinal lesions. The outcome shows that EffTNet is effective in DR severity grading with high accuracy and good interpretability that shows good generalization performance across the two datasets and the possibility for deploying to a large-scale retinal screening system.