多特征融合的超像素谱聚类MRI脑肿瘤图像分割

Segmentation of MRI brain tumor image of superpixel spectral clustering based on multi-feature fusion

为了有效地提高MRI脑肿瘤图像的分割精度,更好地辅助医生诊断病情,提出了一种多特征融合的超像素谱聚类MRI脑肿瘤图像分割方法。首先通过简单线性迭代聚类分割的超像素替代像素点来构建加权无向图,并且融合多种图像特征构建相似度计算函数,同时采用自适应的方式计算高斯核的尺度参数,根据相似度函数计算相似度矩阵进而求得拉普拉斯矩阵,然后对此拉普拉斯矩阵的特征向量进行K-means聚类来完成对图像的分割。在BraTS 2015数据集上与其他2种谱聚类图像分割方法进行了对比实验,并采用相似性系数(Dice)、相对体积误差(RVD)和灵敏度(Sensitivity)这3个指标对分割结果进行评价。结果表明,本文方法在这3个指标上均优于对比方法。因此本文提出的多特征融合的MRI脑肿瘤图像分割方法能够更高效、更精确地完成MRI图像的分割。

In order to effectively improve the segmentation accuracy of MRI brain tumor images and better assist doctors in diagnosis,a segmentation method of MRI brain tumor image of superpixel spectral clustering is proposed in this paper based on multi-feature fusion.Firstly,superpixels obtained by simple linear iterative clustering(SLIC)segmentation are used to replace pixels to construct weighted undirected graph,and similarity calculation function is built based on multi-feature fusion.Meanwhile,the scale parmeters of the Gaussian kernel are calculated by the adaptive computation method,and the similarity matrix is calculated according to the similarity function.Then,the Laplacian matrix is obtained,and the image segmentation is completed by the K-means clustering of the eigen vector of the Laplacian matrix.Finally,this method is compared with the other two spectral clustering image segmentation methods on BraTS 2015 data set,and three indicators such as similarity coefficient(Dice),relative volume deviation(RVD),and sensitivity are used to evaluate the segmentation result.The results show that the method proposed in this paper is better than the contrast methods with respect to these three indicators,therefore,the segmentation method of MRI brain images based on multi-feature fusion can be more robust and more accurate.