基于卷积神经网络的鼻咽肿瘤MR图像分割  

Segmentation of nasopharyngeal Nneoplasm MR images based on convolutional neural network

作　　者：胡光亮 王艳 罗勇[3] 周激流[2] HU Guanghang;WANG Yan;LUO Yong;ZHOU Jiliu(College of Electronics and Information Engineering,Sichnan University,Chengdu Sichnan 610065,China;College of Computer Science,Sichnan University,Chengdu Sichnan 610065,China;Department of Oncology,West China Hospital of Sichnan University,Chengdu Sichnan 610041,China)

机构地区：[1]四川大学电子信息学院,成都610065 [2]四川大学计算机学院,成都610065 [3]四川大学华西医院肿瘤科,成都610041

出　　处：《计算机应用》2018年第A01期208-212,共5页journal of Computer Applications

基　　金：国家自然科学基金资助项目(61701324)

摘　　要：鼻咽肿瘤区域形状多变、边界模糊,利用专家手工勾画确定放疗靶区不仅耗时,而且还会带有一定的主观性。针对这一问题,提出一种利用卷积神经网络(CNN)对鼻咽肿瘤MR图像全自动的分割方法。首先,将每一个鼻咽肿瘤MR图像分成大小为32×32的图像块;其次,用图像块训练CNN结构来实现对鼻咽肿瘤的全自动分割。CNN结构中卷积核的大小均为3×3,采用小尺寸的卷积核能够提取出更加丰富的局部特征,并且能够设计出更深的网络结构。为了验证CNN结构的性能,从测试样本集中随机选取了60个鼻咽肿瘤切片进行分割,经过形态学处理后最终的分割精度可达到:Dice系数83. 39%,PM系数82. 10%,CR系数74. 47%。实验结果表明,该CNN分割方法能够很好地分割出鼻咽肿瘤区域。Nasopharyngeal neoplasms have high variations in shape and size, as well as blurred boundary. Clinicians manually delineate the target area of radiotherapy, which is not only time-consuming, but also with some subjectivity. In view of this problem, an automatic segmentation method of nasopha1：cngeal neoplasm MR images based on Convolutional Neural Network （CNN） was proposed. Firstly, each nasopharyngeal neoplasm MR image was divided into 32 × 32 patches; Then, the CNN structure was trained with the patches to realize the automatic segmentation of nasopharyngeal neoplasms. The size of the convolution kernels in the CNN structure was 3 × 3, small kernels were used to extract more rich local features, also allows designing a deeper network structure. To evaluate the performance of the proposed CNN, sixty nasopharyngeal neoplasm slices were randomly selected from the test dataset for segmentation. After the morphological processing, the final segmentation results are： Dice coefficient 83.39%, PM coefficient 82.10%, CR coefficient 74.47%. The experimental results show that the CNN segmentation method proposed in this paper can effectively segment the nasopharyngeal neoplasms.

关 键 词：卷积神经网络 鼻咽肿瘤 MR图像 医学图像分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]