Abstract:In fundus disease classification, subtle pathological features are difficult to extract, and scarcity of high-quality annotated fundus images easily leads to model overfitting. To address these problems, a lightweight fundus disease classification network based on cross-fusion of multi-scale channel attention (MSCA) and Manhattan spatial decay attention (MSDA) is proposed. To accurately extract fine-grained features, a cross-fusion attention (CFA) mechanism is designed. MSCA and MSDA are alternately fused through convolution, which enhances sensitivity to subtle fundus lesions. On this basis, a hierarchical selective attention insertion strategy is adopted, in which attention mechanisms are allocated according to the feature requirements of different network layers, enabling effective extraction of multi-level fundus image features. In addition, to alleviate overfitting caused by data scarcity, an importance-based pruning strategy combining attention scores and norm evaluation is applied. Redundant modules are removed to reduce model parameters and avoid excessive extraction of irrelevant features and background noise. Experimental results show that the proposed network achieves an improvement of 5.07–12.68 percentage points in classification accuracy and demonstrates significant advantages in feature extraction capability and lightweight design.