Multimodal architecture based on dual attention improvement to assist in glaucoma grading challenges

Diabetic glaucoma, a prevalent complication of diabetes, seriously threatens patients' visual health and ranks as a leading cause of irreversible blindness among diabetic individuals. In clinical practice, accurate grading of diabetic glaucoma is crucial for formulating personalized treatment strategies and predicting disease progression. However, existing methods often rely on single - modality data or lack effective integration of multi-modal information, leading to suboptimal grading performance. This study aims to address these issues by presenting two multi-modal disease grading networks enhanced by the cross-attention mechanism for grading diabetic glaucoma using OCT and color fundus images. Two novel cross-attention based image fusion strategies are developed: one employs the multi - head cross - attention mechanism for better inter-modal information fusion, and the other combines self-attention and cross-attention mechanisms. The experimental results show that using the proposed cross attention based method, the Kappa value of multimodal scoring in the GAMMA challenge reaches 84.4%, and the F1 score is 81%. Our work exceeded the performance of the champion and runner up network models in that year's competition. Also, multi-modal grading accuracy shows a 1% - 4% increase compared to single-modality grading. This research not only improves the accuracy of diabetic glaucoma grading but also provides valuable feature support for future multi-task learning models.