基于词袋模型的脑肿瘤MR图像分割方法  

Segmentation of brain tumor for MR images based on bag of words

作　　者：赵建奇[1] 黄美艳[1] 冯前进[1] 陈武凡[1] 

机构地区：[1]南方医科大学生物医学工程学院,广东广州510515

出　　处：《计算机工程与设计》2014年第4期1312-1317,共6页Computer Engineering and Design

基　　金：国家863高技术研究发展计划基金项目(2012AA02A616)

摘　　要：提出了一种基于词袋模型配合滑动窗口提取像素点特征的脑肿瘤MR图像分割方法。通过预处理去除非脑组织并进行灰度值归一化,采集图像的图像块特征并聚类生成视觉词典;在生成视觉词典的过程中,通过分别生成病灶区域词典和背景区域词典进而组合得到最终的联合视觉词典;利用联合词典配合滑动窗口对像素点进行表达并将其作为像素点的特征,利用逻辑回归分类器进行训练和分类从而完成对脑肿瘤的分割。在160幅脑肿瘤MR图像组成的数据集上进行实验,实验结果表明分割准确率达到90.42%。A new method for segmentation of brain tumor in MR images using bag of words model along with sliding window is proposed. Firstly, preprocessing is performed on dataset to get rid of the non-brain tissue and normalize the grey value. Second- ly, the local patch descriptor is extracted from the images and the visual vocabulary is constructed by clustering algorithms. When constructing the visual vocabulary, the vocabulary of lesion and the vocabulary of background are constructed respectively, the two category-specific vocabularies are combined to form a union visual vocabulary. The bag of visual words representation of pixel is treated as the pixel＇s feature. Finally, a logistic regression classifier is used for training and classifying, finishing the seg- mentation of brain tumor. The method is evaluated in a T1 MR image dataset comprised of 160 patients and the results show that the segmentation accuracy can achieve 90.42%.

关 键 词：MR图像 脑肿瘤分割 词袋模型 分类词典 滑动窗口 分类器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]