Frequency—Semantic Dual—Driven Segmentation of Peripapillary Atrophy in Retinal Fundus Images

Peripapillary atrophy (PPA) is a critical imaging biomarker for the diagnosis and grading of pathological myopia (PM), and its accurate segmentation plays an essential role in clinical evaluation and longitudinal monitoring. However, automatic PPA segmentation from color fundus images remains challenging due to blurred and discontinuous boundaries, low tissue contrast, and noise interference, which limit the performance of existing methods. In this study, we propose a novel segmentation framework, termed FSD-PPA (Frequency-Semantic Dual-Driven Segmentation Network), to address these limitations. The proposed FSD-PPA integrates a high-frequency edge extraction module to enhance anatomical boundary perception, a low-frequency compensation module to restore tissue structural integrity, and a multi-scale semantic enhancement module to improve cross-level contextual understanding. Experimental results on an expert-annotated PPA dataset and the public PALM dataset demonstrate that our method achieves Dice scores of 83.48 % and 84.30 %, respectively, outperforming existing approaches. These results confirm two major advantages of FSD-PPA: (1) high segmentation accuracy, enabling precise identification of complex PPA morphology; and (2) strong generalization ability, with stable performance across different datasets. This study provides a robust and reliable solution for automatic PPA segmentation. Clinically, it holds potential as an assistive tool for early PM screening and offers an objective imaging basis for quantitative assessment and monitoring of myopic progression, thereby facilitating intelligent diagnosis of myopia-associated fundus lesions.lligent diagnosis systems for myopia-related fundus lesions.