Glaucoma classification with MBConv and Channel Attention Network

Glaucoma is a leading cause of optic nerve damage and irreversible blindness. Existing single-modality-based diagnostic methods are difficult to comprehensively capture the characteristic data of glaucoma, which is prone to often lead to missed and misdiagnosed diagnoses. To address the above issues, a multimodal glaucoma grading method called MCANet is proposed in this paper. This method develops a dual-branch neural network structure for precise glaucoma classification by combining 2D fundus images and 3D OCT images. Among them, the fundus image branch is based on EfficientNet-B3. The OCT image branch proposes the MSBConv module, which is based on ResNet34 for modular reconstruction and integrates MBConv and SE attention mechanisms. This enhances the ability to identify structural indicators. In addition, this paper introduces a Hybrid Loss function based on Cross-Entropy. By weighted fusion of cross entropy and mean square error, this model effectively maintains the structural continuity in OCT images while improving the multi-classification accuracy. The performance of MCANet and other models in glaucoma grading tasks was validated on the GAMMA multimodal glaucoma dataset. The results indicate that MCANet has an accuracy of 94.3% and a Kappa value of 76.7, both of which are superior to other models.