A multi-disease classification method for fundus images integrating gated central pixel-difference convolution and semantic knowledge alignment

Multi-disease classification of fundus images is a critical task in intelligent ophthalmological screening; however, the visual manifestations of different diseases vary substantially, and existing methods still insufficiently exploit local structural variations and disease-specific semantic priors. To address this issue, this paper proposes a multi-disease classification method for fundus images that integrates gated central pixel-difference convolution with semantic knowledge alignment. In the shallow feature extraction stage, the proposed method introduces gated central pixel-difference convolution (gPDC-C), which employs a gating mechanism to fuse central difference responses with intensity responses, thereby enhancing shallow local structural features. In the deep feature extraction stage, a semantic knowledge alignment (SKA) module is designed to incorporate disease-specific semantic priors into visual feature learning, strengthening the consistency between deep visual features and categorical semantic prototypes. On the Eye Diseases Classification (EDC) dataset, SKA-gPDCNet achieves accuracy, sensitivity, F1-score, area under the receiver operating characteristic curve (AUC), and specificity values of 93.73%, 93.62%, 93.64%, 98.99%, and 97.92%, respectively. Under a unified protocol, progressive ablation experiments showed modest mean-level gains over the ResNet34 baseline and supported the complementary contributions of gPDC-C and SKA. After retraining on ODIR, SKA-gPDCNet achieved mean F1-scores of 66.91% and 70.46% and mean sensitivities of 66.67% and 71.82% on the official Off-site and On-site test splits, respectively; however, direct EDC-to-ODIR transfer yielded low absolute performance, indicating substantial domain shift and limited cross-domain transferability.