Segmentaion of Parapapillary Atrophy in Retinal Images using HED

Abstract

Parapapillary atrophy (PPA) is a kind of atrophic abnormality of eyes, which lies on the retinal fundus adjacent to optic disc (OD). Clinically, PPA is correlated with several pathologies of eyes. For example, the area of PPA has a strong association with the severity of myopia. In this work, we developed a new segmentation algorithm aimed at PPA in retinal images. The frame work was derived from Holistically-Nested Edge Detection (HED), which originally deals with edge detection. Firstly, extracting region of interest (ROI) and normalizing datasets were performed in preprocessing stage. Next, an end-to-end model began to acquire rich hierarchical features with deep supervision on side outputs. We applied a new balance parameter to loss function for realization of the area segmentation. Finally, the holistic image predictions (PPA area segmented) were performed through full convolutional neural layers. The model training on retinal images with annotation from professional ophthalmologist achieves great performance on both adults and pupils datasets. Experiential results present the average F-score of 0.7910 on adults test data (200 images) and 0.7478 on pupils (50 images). MIoU of the segmentation reaches 0.7237 of adults and 0.6892 of pupils.