Extracting Retinal Anatomy and Pathological Structure Using Multiscale Segmentation

Abstract

Fundus image segmentation technology has always been an important tool in the medical imaging field. Recent studies have validated that deep learning techniques can effectively segment retinal anatomy and determine pathological structure in retinal fundus photographs. However, several groups of image segmentation methods used in medical imaging only provide a single retinopathic feature (e.g., roth spots and exudates). In this paper, we propose a more accurate and clinically oriented framework for the segmentation of fundus images from end-to-end input. We design a four-path multiscale input network structure that learns network features and finds overall characteristics via our network. Our network’s structure is not limited by segmentation of single retinopathic features. Our method is suitable for exudates, roth spots, blood vessels, and optic discs segmentation. The structure has general applicability to many fundus models; therefore, we use our own dataset for training. In cooperation with hospitals and board-certified ophthalmologists, the proposed framework is validated on retinal images from large databases and can improve diagnostic performance compared to state-of-the-art methods that use smaller databases for training. The proposed framework detects blood vessels with an accuracy of 0.927, which is comparable to exudate accuracy (0.939) and roth spot accuracy (0.904), providing ophthalmologists with a practical diagnostic and a robust analytical tool.