MPMNet: Modal Prior Mutual-Support Network for Age-Related Macular Degeneration Classification

Early screening and classification of Age-related Macular Degeneration (AMD) are crucial for precise clinical treatment. Currently, most automated methods focus solely on dry and wet AMD classification. However, the classification of wet AMD into more explicit type 1 choroidal neovascularization (CNV) and type 2 CNV has rarely been explored, despite its significance in intravitreal injection. Furthermore, previous methods predominantly utilized single-modal images for distinguishing AMD types, while multi-modal images can provide a more comprehensive representation of pathological changes for accurate diagnosis. In this paper, we propose a Modal Prior Mutual-support Network (MPMNet), which for the first time combines OCTA images and OCT sequences for the classification of normal, dry AMD, type 1 CNV, and type 2 CNV. Specifically, we first employ a multi-branch encoder to extract modality-specific features. A novel modal prior mutual-support mechanism is proposed, which determines the primary and auxiliary modalities based on the sensitivity of different modalities to lesions and makes joint decisions. In this mechanism, a distillation loss is employed to enforce the consistency between single-modal decisions and joint decisions. It can facilitate networks to focus on specific pathological information within individual modalities. Furthermore, we propose a mutual information-guided feature dynamic adjustment strategy. This strategy adjusts the channel weights of the two modalities by computing the mutual information between OCTA and OCT, thereby mitigating the influence of low-quality modal features on the network's robustness. Experiments on private and public datasets have demonstrated that the proposed MPMNet outperforms existing state-of-the-art methods.