Gaussian Distribution Prior Based Multi-view Self-supervised Learning for Serous Retinal Detachment Segmentation

Abstract

Assessment of serous retinal detachment (SRD) plays an important role in the diagnosis of central serous chorioretinopathy (CSC). In this paper, we propose an unsupervised method, called Gaussian distribution prior based Multi-view Self-supervised Learning (G-MSL), for the segmentation of SRD, in spectral domain optical coherence tomography (SD-OCT) images. We firstly count the Gaussian distribution prior for each targeted retinal layer from normal SD-OCT images. Then the Gaussian distribution prior-based fitting detects the abnormal pixels belonging to SRD in each targeted retinal layer. The generated coarse SRD region masks are used for self-supervised learning to optimize the SRD regions. The fully connected conditional random field is applied to obtain the SRD segmentation results. To improve the robustness of the proposed method for 3D SD-OCT volumes, we repeatedly carry out the above-mentioned operations from another view. The final segmentation results are obtained by getting the union of the results of multiple views. Experimental results on 20 subjects with CSC demonstrate that the proposed method can achieve the average dice similarity coefficient of 91.69%. G-MSL shows enough potential for the improvements of the clinical CSC evaluation and achieves higher segmentation accuracy than the existing supervised deep learning methods when the training set is not very large.

This study was supported in part by National Natural Science Foundation of China (62172223, 61671242), in part by Key R&D Program of Jiangsu Science and Technology Department (BE2018131) and the Fundamental Research Funds for the Central Universities (30921013105).