Automated diagnosis of pathological myopia based on fundus color photography and dynamic multi-scale learning

Pathological myopia (PM) and high myopia have emerged as leading causes of irreversible visual impairment and blindness worldwide. In clinical practice, the significant variability in lesion scales within fundus color photography (FCP) poses a substantial challenge for automated systems. Traditional single-scale deep learning models often struggle to capture heterogeneous pathological features, thereby limiting diagnostic precision. In this study, we propose a novel Task-Aware Dynamic Multi-Scale Network (TADMS-Net) specifically designed for myopia grading and PM identification using FCP. The architecture integrates a Dynamic Multi-Scale Pyramid (DMSP) module to enable the precise extraction and adaptive enhancement of multi-scale lesion characteristics. To optimize performance, a Token-Free Transformer (TFT) module is implemented to capture long-range global dependencies while significantly reducing computational overhead. Furthermore, we introduce the Task-Guided Global-Local Hierarchical Attention (TGGLHA) module, which facilitates the global localization of suspicious lesion areas followed by rigorous refinement of local pathological details. To ensure clinical interpretability, Grad-CAM is employed to generate visual heatmaps, highlighting the critical retinal regions that drive the models diagnostic decisions. Extensive experiments conducted on the MG-PM dataset and two benchmark datasets (iChallenge-PM and Kaggle EyePACS) demonstrate that TADMS-Net consistently outperforms seven state-of-the-art baseline models. TADMS-Net offers a robust and interpretable solution for accurate and efficient automated myopia grading and pathological myopia detection in clinical fundus imaging.