基于3D卷积神经网络的脑肿瘤医学图像分割优化

曹祺炜 王峰 牛锦

摘  要： 针对2D卷积神级网络不能够较好地提取各模态之间的差异信息，不同的图像层肿瘤大小差异显著，且分割精度低，单模态MRI无法清晰地反映GBM的不同组织结构，提出一种基于3D多池化卷积神经网络拟解决以上实际问题。将卷积神经网络应用到脑肿瘤分割上，并针对脑肿瘤的特点，提出3D多池化卷积神经网络模型，通过多尺度的输入与多尺度的下采样，且在后端使用条件随机场（CRF）使图片尽量在边界处分割，增加图像的分割精度，克服脑肿瘤的个体差异，同时适应脑肿瘤不同图像层之间的大小位置差异。通过对100例患者的多模态磁共振图像进行分割，Dice系数达到91.64%;MRI脑肿瘤分割的改进方法可使分割精度得到明显提高，可更好地提取各模态之间的差异信息，实现适应范围更广的MRI肿瘤分割，并准确有效地分割脑肿瘤。

关键词： 脑肿瘤; 医学图像分割; 多模态MRI; 差异信息提取; 多尺度采样; 3D卷积神经网络

中图分类号： TN911.73?34; TP399                    文献标识码： A                 文章编号： 1004?373X（2020）03?0074?04

Optimization of brain tumor medical image segmentation

based on 3D convolutional neural network

CAO Qiwei， WANG Feng， NIU Jin

（School of Information and Computer Science， Taiyuan University of Technology， Taiyuan 030024， China）

Abstract： A 3D?based multi?pooling convolutional neural network is proposed in view of the facts that the 2D convolutional neural network can′t extract the difference information among different modes with satisfaction， the differences of tumor size for different image layers are obvious， the image segmentation accuracy is low， and the single?mode MRI （magnetic resonance imaging） fails to clearly reflect the different tissue structures of GBM （glioblastoma multiforme）. The convolutional neural network is applied to brain tumor segmentation， and a 3D multi?pooling convolutional neural network model is proposed according to the characteristics of brain tumors. In the proposed model， multi?scale input and multi?scale down?sampling are implemented and CRF（conditional random field） is used at the rear?end to segment the image at the boundary as far as possible， so as to improve the image segmentation accuracy， eliminate the individual differences of brain tumor images and adapt to the size and position differences in different image layers of brain tumors. The Dice coefficient reached 91.64% by segmenting the multi?mode magnetic resonance images of 100 patients. The improved MRI brain tumor segmentation method can significantly improve the segmentation accuracy， better extract the difference information among various modes， adapt to a wider range of MRI tumor segmentation， and accurately and effectively segment brain tumors.

Keywords： brain tumor; medical image segmentation; multi?mode MRI; difference information extraction; multi?scale sampling; 3D convolutional neural network

0  引  言

医学图像中病灶的分割和定量评估为神经病理学的分析提供了有价值的信息，在规划治疗策略，监测疾病进展和预测患者结果中扮演重要角色。因此，图像的分割效率和准确性在医学病例分析、临床诊断以及治疗方面具有重要意义。本文基于2D多池化卷积神经网络基础，采用数学模型和高维重建方法对2D?CNNs（2D Convolutional Neural Networks）进行改进，提出基于3D多池化卷积神经网络解决以上实际问题。

本文提出一种多模态多池化3D卷积神经网络脑肿瘤图像分割方法。……