Multi-view diabetic retinopathy grading via cross-view spatial alignment and adaptive vessel reinforcing

Our research introduces a novel deep learning framework that leverages multi-view fundus images for Diabetic Retinopathy (DR) grading. Existing models for fundus image analysis often prioritize salient features, such as the optic disk, potentially overlooking finer details critical for DR detection, like retinal vessel information. To address this, we introduce a learnable retinal vessel reinforcement block to enhance the representation of retinal vessels. Additionally, recognizing the limitations of traditional multi-view models in capturing the spatial correlation between 2D appearances from different views, we propose a cross-view spatial region aligning vision transformer (ViT). This ViT-structured model is crucial for modeling cross-view relationships and integrating lesion information across individual views. Furthermore, a multi-view decision fusion module synergistically fuses diagnostic insights from multiple perspectives, enhancing the model's diagnostic capabilities. Our method demonstrates significant superiority over existing single-view and multi-view models across key performance metrics, including accuracy, precision, sensitivity, specificity, and F1 score.