Diagnosing Pathologic Myopia by Identifying Posterior Staphyloma and Myopic Maculopathy Using Ultra-Widefield Images with Deep Learning

Pathologic myopia (PM) has long been a leading cause of visual impairment and blindness. While numerous deep learning-based approaches have improved the efficiency and accuracy of recognizing PM, few have thoroughly investigated clinically significant pathological patterns due to the scarcity of datasets with lesion-wise labeling, particularly those comprising ultra-widefield (UWF) images that encompass a broader retinal field of view. In this study, we gather a large-scale multi-source ultra-widefield imaging myopia dataset, PSMM, labeled with posterior staphyloma (PS) and myopic maculopathy (MM). Compared with traditional colored fundus photography, UWF images exhibit informative characteristics concerning peripheral lesions caused by axial elongation and structural deformation in eyes with pathologic myopia. The labels obtained from the dataset can substantially assist in the progression diagnosis of pathologic myopia and guide prognosis. We introduce an end-to-end lightweight framework called RealMNet, which precisely identifies these challenging pathological patterns underpinned by a well-curated dataset. RealMNet is more adaptable to medical devices with only 21 million parameters compared to existing approaches. Through extensive experiments on a unified platform using all-around metrics regarding bipartitions and rankings across three experimental protocols, we demonstrate the robustness and generalizability of RealMNet, showcasing promising merit in clinical applications.