Pathological myopia classification with simultaneous lesion segmentation using deep learning

Pathological myopia (PM) is the seventh leading cause of blindness, with a reported global prevalence up to 3%. Early and automated PM detection from fundus images could aid prevent blindness in world population that characterized by rising prevalence. We aim assess use convolutional neural networks (CNNs) for semantic segmentation myopia-induced lesions on recently introduced reference data set. This investigation reports results CNNs developed Myopia (PALM) dataset, which consists 1200 images. Our CNN bundles lesion classification, as two tasks are heavily intertwined. Domain knowledge also inserted through introduction new Optic Nerve Head (ONH)-based prediction enhancement atrophy fovea localization. Finally, we first approach localization using instead or regression models. Evaluation metrics include area under receiver operating characteristic curve (AUC) detection, Euclidean distance localization, Dice F1 (optic disc, retinal detachment). Models trained 400 available training achieved an AUC 0.9867 58.27 pixels task, evaluated test set scored 0.9303 0.9869 optic 0.8001 0.9135 atrophy, 0.8073 0.7059 detachment, respectively. report successful simultaneous classification pathological associated lesions. work was acknowledged award context “Pathological images” challenge held during IEEE International Symposium Biomedical Imaging (April 2019). Considering (pathological) cases often identified false positives negatives glaucoma deep learning models, envisage current future research discriminate between glaucomatous highly-myopic eyes, complemented landmarks such fovea, disc atrophy.