Refinement of Parapapillary Atrophy Segmentation Based on Conditional Random Field

Abstract

Parapapillary atrophy (PPA) is a crescent atrophic abnormality adjacent to optic disk (OD). In clinical findings, PPA is highly related to glaucoma and high myopia, which makes it a potential indicator for ophthalmic diseases. Therefore, it is significant to develop an algorithm for automatic segmentation of PPA so as to aid the treatment and the prevention of eye diseases. A method that combines the confidence map with conditional random field (CRF) model is proposed in this paper to achieve the optimization of initial segmentation results from neural network. Holistically-Nested Edge Detection (HED) is applied to the segmentation of PPA and a confidence map is obtained. The CRF model is constructed with the confidence map and Maximum a posteriori (MAP) inference is implemented to conduct the segmentation. Post-processing is utilized to further refine the segmentation result. The proposed method has the advantage of robust performance when there is difference between training set and test set. The experimental results show that the average F-score is increased from 0.57 to 0.67, which shows a fair improvement effect.