Automated analysis of fundus images plays a crucial role in ophthalmological disease diagnosis. However, existing research faces three major challenges: First, most methods either focus on single-eye image analysis or use simple binocular feature fusion strategies, failing to fully exploit the complementarity of binocular information. Second, single-scale feature extraction strategies struggle to capture the multi-scale features of various fundus lesions effectively. Finally, existing methods typically model single diseases, unable to accurately characterize dynamic relationships between diseases in multi-label tasks. To address these challenges, we propose a multi-scale binocular fusion semantic co-occurrence network (MBFSNet). This method first uses a high-resolution network (HRNet) as the backbone to extract multi-scale feature representations, then achieves dynamic fusion of binocular features through the multi-scale binocular pixel-wise feature fusion module (MB-PFFM). It effectively integrates features from different scales using the cross-scale feature aggregation module (CFAM) and constructs a disease semantic relation graph (DSRG) to dynamically model complex dependencies between disease categories. The proposed method achieves improvements of 1.2% and 2% in final score compared to existing approaches on the off-site and on-site test sets of the OIA-ODIR dataset, respectively, providing a new solution for automated and precise ophthalmological disease screening.