一种基于CA-Net的脑肿瘤磁共振图像分割算法  

A method of brain tumor segmentation based on CA-Net

作　　者：柯盼盼 陈胜[1] 李珂然 KE Panpan;CHEN Sheng;LI Keran(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《智能计算机与应用》2022年第9期208-213,F0003,共7页Intelligent Computer and Applications

基　　金：国家自然科学基金(81101116)。

摘　　要：磁共振影像是脑肿瘤疾病中常用的诊断工具,临床上的量化分析需要对影像结果进行分割得到肿瘤区域,但手动分割十分耗时且高度依赖于医生的临床经验。为此,本文提出一种基于深度学习的脑磁共振图像分割算法,解决了原网络无法有效提取关键特征的问题。该算法使用改进的有限对比度自适应直方图均衡(Contrast Limited Adaptive Histogram Equalization, CHALE)算法,对脑肿瘤磁共振影像进行图像增强后,将结果输入CA-Net网络对数据集初步分割,并将全注意力算法和U-net骨架结构结合(包括空间、通道和尺度注意力模块),实现对不同尺度的空间和通道的特征转换连接。模型应用混合损失函数提高分割精度。初步分割的结果可通过后处理进一步提高精度,得到最终的肿瘤区域。初步分割结果中,Dice指标可以达到88.40(±0.24)%,结合图像处理提高至89.21(±0.36)%,分割精度相较于其它算法有明显提高。Magnetic resonance imaging is a common diagnostic tool in brain tumor diseases. Clinical quantitative analysis needs to segment the image results to obtain the tumor area, but manual segmentation is very time-consuming and highly dependent on the clinical experience of the doctor. Therefore, this paper proposes a brain magnetic resonance image segmentation algorithm based on deep learning. Firstly, the algorithm enhances the magnetic resonance image of brain tumor using improved contrast limited adaptive histogram equalization(CHALE). The obtained results are input into CA-Net network to preliminarily segment the dataset. The network combines the full attention algorithm with the U-net skeleton structure, including spatial, channel and scale attention module to realize the feature conversion and connection of different scales of the space and channels. The model uses mixed loss function to improve the segmentation accuracy. The preliminary segmentation results can further improve the accuracy through post-processing to obtain the final tumor region. In the preliminary segmentation results, the Dice index can reach 88.40(±0.24)%, which can be improved to 89.21(±0.36)% combined with image processing. The segmentation accuracy is significantly improved compared with other algorithms.

关 键 词：图像分割 深度学习 CA-Net 注意力模块 混合损失函数 后处理 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]