Abstract

Info

Title:: Semantic segmentation of ultrastructural pathological images of glomerular filtration membrane using deep learning network

Author(s):: WEN Jiayuan1; LIN Guoyu1; ZHANG Yiwen1; ZHOU Zhitao2; CAO Lei1; FENG Qinchang3; 1. School of Biomedical Engineering, Southern Medical University, Guangzhou 510515, China; 2. Electron Microscope Room, Central Laboratory, Southern Medical University, Guangzhou 510515, China; 3. Guangdong Institute of Medical Instruments, Guangzhou 510500, China

Keywords:: Keywords: deep learning; DeepLab; glomerular filtration membrane; ultrastructural pathological image; semantic segmentation

Abstract:: Abstract: The glomerular filtration membrane contains 3 layers of ultrastructures, namely endothelial cells, glomerular basement membrane and podocytes. The morphological change of glomerular filtration membrane is one of the important indicators for the diagnosis of glomerular diseases. The accurate semantic segmentation of filtration membrane helps pathologists identify and determine the subtle pathological changes in filtration membrane so as to provide reference in the related pathological diagnosis. Due to the complicated structure and low gray-scale resolution of the ultrastructural pathology image of glomerular filtration membrane, the traditional segmentation algorithms can only segment the basement membrane with the simplest morphology, and it is difficult to guarantee the segmentation accuracy. Herein an automatic semantic segmentation algorithm based on deep learning network DeepLab-v3 for glomerular filtration membrane is proposed. Atrous convolution is used to expand field-of-views and control the feature resolution of the image; and then multi-scale image information is obtained through atrous spatial pyramid pooling; and finally, the 3 components of glomerular filtration membrane are simultaneously segmented, and all can achieve a favorable segmentation effect. Through the experimental exploration on important parameters, the average segmentation accuracy can reach 0.776, and a relatively good model at present is obtained in the study.