Glioma Image Segmentation Method on Fully Convolutional Neural Network

Abstract

Aiming at the difference in the segmentation performance of the three segmentation target regions in the glioma image segmentation task based on the fully convolutional neural network, we propose a comprehensive evaluation method of neural network performance based on four evaluation indices. In addition, we analyze the performance and characteristics of neural network in the segmentation task of glioma, study the segmentation performance of neural network in the whole tumor (WT), tumor core (TC) and enhanced tumor (ET) regions, and propose a deep learning algorithm based on multiple networks in parallel. In this paper, the input image of the two-dimensional neural network is sliced, and the input of the three-dimensional neural network is processed in two ways: overlapping and non-overlapping, and in the image post-processing part, the three-dimensional image is reconstructed before the evaluation index is calculated. This article uses four evaluation indexes, which are Dice, Sensitivity, PPV, and Hausdorff, for the three segmentation target regions, and performs RSR* weight calculation, and finally performs a comprehensive evaluation. Experimental results show that Vnet has the best comprehensive segmentation performance, FCN-8s has the best segmentation performance in the TC area, Unet++ has the best segmentation performance in the ET area, and Vnet has the best segmentation performance in the WT area. Based on this, we propose a FUV multi-network parallel algorithm, combined with a reverse attention mechanism to improve the segmentation accuracy of the three segmentation target regions.