Bi-Directional Hybrid Attention Feature Pyramid Network for Detecting Diabetic Macular Edema in Retinal Fundus Images

Abstract

Diabetic macular edema (DME) is an advanced stage of Diabetic Retinopathy (DR), caused when exudates encroach on the macula region of the retina. Early detection of DME symptoms can avert vision impairment in most of the cases. A comprehensive and automated detection framework will greatly facilitate early detection of DME in diabetic patients. In this work, we propose a convolutional neural network based model, entitled as Bi-directional Hybrid Attention based Feature Pyramid Network (BhAFPN), to grade diabetic macular edema stages. The proposed model uses a bi-directional fusion of multi-scale features, augmented through hybrid-attention mechanisms implanted in the bottom-up and top-down feedback paths. A multi-level global attention block (GAB) mechanism weighs the augmented feature maps obtained from different scales according to their contribution to the final DME grade prediction. The model includes SVM classifier to perform DME screening from the extracted features of the fully connected layer (combinedly referred as BhAFPN+SVM). In addition, the network is also able to localize suspicious DME anomaly regions in fundus images in a weakly-supervised manner. The proposed model is significantly more light-weighted than the state-of-the-art (SOTA) models in DME grading task. The generalization of the proposed model is tested on multiple public retinal datasets. The proposed model outperformed all the reviewed SOTA methods on IDRiD and MESSIDOR-1 datasets in DME severity grading task by achieving the highest AUC-ROC scores of 0.9901 and 0.9946, respectively. Likewise, the BhAFPN+SVM model outperformed the SOTA models in DME screening task on both MESSIDOR-1 and MESSIDOR-2 datasets by achieving the highest accuracy of 95.92% and 90.83%, respectively. The interpretability of the model is analyzed using Grad-CAM-based heatmap visualization method and t-SNE plots. The lightweight nature of the proposed model along with its effective learning capabilities and robust performance in DME grading task makes it a viable and optimal candidate to be integrated with a real-time multi-disease detection systems.