Hierarchical Multi-Scale Synergistic Fusion Network for Pathological Myopia ATN Classification

Deep learning has achieved groundbreaking progress in the field of image recognition, yet existing networks have not fully explored the potential connections between network architecture and fundus image features. To address this issue, this paper proposes the Hierarchical Multi-Scale Synergistic Fusion Network (HMSFNet), aimed at resolving the problems of feature collapse in CNNs and the performance limitations of ensemble learning fusion mechanisms. HMSFNet is based on a dual-branch integrated architecture that analyzes key features in fundus images from multiple dimensions and embeds a feature synergy stability module to promote mutual learning between the two branches. Additionally, this study designs a novel and more efficient attention mechanism focused on multi-scale feature recognition of lesion areas. The study constructed a dataset comprising 11,495 high-quality fundus images. In benchmark tests, HMSFNet achieved optimal performance in diagnosing atrophic lesions (A), tractional lesions (T), and neovascular lesions (N), with accuracies of 94.9%, 95.82%, and 99.4%, respectively, significantly outperforming existing models. This achievement not only demonstrates the effectiveness of HMSFNet in diagnosing pathological myopic maculopathy (ATN) but also provides a powerful auxiliary tool for clinical applications.