融合密集连接与多尺度卷积的脑肿瘤MRI图像分割算法  

Brain Tumor MRI Image Segmentation Algorithm Fused with Dense Connections and Multi-Scale Convolution

作　　者：杨述斌[1] 王锋 董春林 YANG Shubin;WANG Feng;DONG Chunlin(School of Electrical and Information Engineering,Wuhan Institute of Technology,Wuhan 430205,China)

机构地区：[1]武汉工程大学电气信息学院,湖北武汉430205

出　　处：《电视技术》2022年第10期32-36,共5页Video Engineering

摘　　要：针对现有脑肿瘤分割算法存在上下层联系匮乏、特征欠提取而导致分割精度低的问题,提出一种融合密集连接与多尺度卷积的U型算法,实现对脑肿瘤病灶的精确分割。通过对传统的U-Net算法进行改进,在编码和解码部分引入多尺度卷积、密集连接以及空洞卷积,增强算法对病灶特征的提取以及定位能力。使用BraTs2019数据集对改进后的算法进行实验验证,脑肿瘤整体区域、核心区域、增强区域的Dice分割指数达到0.8532,0.8721,0.7960,对应的Hausdorff距离分别为2.6089,1.5852,2.7416,相较于其他算法均有不同程度的提升。实验结果表明,所提算法可有效地实现对脑肿瘤的精准分割。To address the problem of low segmentation accuracy due to the lack of upper and lower layer connections and underextraction of features in existing brain tumour segmentation algorithms,this paper proposes a U-shaped algorithm fusing dense connectivity and multi-scale convolution to achieve accurate segmentation of brain tumour lesions.Enhance feature extraction and localization by improving the traditional U-Net algorithm by introducing multi-scale convolution,dense concatenation and null convolution in the encoding and decoding sections.Using the BraTs2019 dataset to experimentally validate the improved algorithm,the Dice segmentation indices of the overall brain tumour region,core region and enhanced region reached 0.8532,0.8721 and 0.7960,corresponding to Hausdorff distance of 2.6089,1.5852 and 2.7416,respectively,which compared to the classical algorithm and some advanced The coefficients are 2.6089,1.5852 and 2.7416 respectively,which are different degrees of improvement compared to the other algorithm.The experimental results show that the proposed algorithm can effectively achieve accurate segmentation of brain tumours.

关 键 词：脑肿瘤分割 U-Net 密集连接 空洞卷积 多尺度卷积 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]