DEEP LEARNING-ENHANCED DETECTION OF PATHOLOGICAL MYOPIA VIA MULTI-FEATURE FUSION AND PRECISE RETINAL SEGMENTATION

Abstract

Pathological Myopia (PM) is a major cause of permanent blindness in the world, but it is difficult to detect during its early development as there is a complex and diverse manifestation of retinal lesions. This research provides an automated diagnostic scheme, which makes use of a multi-feature fusion model to categorize fundus images into normal, highly myopic and pathologically myopic groups. The system identifies five clinical biomarkers, including Fuchs Spots, Lattice Degeneration, Posterior Staphyloma, Tigroid Fundus, and Peripapillary Atrophy (PPA). To combat the problem of inaccurate localization of optic disk, one of the main errors in the previous PPA-based models, this work uses a U-Net deep learning structure to perform high-quality segmentation. A combination of the Gray Level Co-occurrence Matrix (GLCM) which was used to analyze texture and Oriented FAST and Rotated BRIEF which was used to detect localized lesions were used to extract features. These multi-dimensional characteristics were fed to Support Vector Machine (SVM) classifier. The combined model was tested on a test set of 400 fundus images and the individual model showed a diagnostic accuracy of about 90.15. The 10-fold cross-validation revealed that the system is robust, as the U-Net-enhanced PPA feature has the highest score of 92.27 in the validation. These results indicate that a multi-feature analysis is a useful way of overcoming the drawbacks of mono-marker detection and can be viewed as a reliable and objective tool to detect myopic blindness and apply preventive strategies in the course of early clinical intervention.