基于LBP和极限学习机的脑部MR图像分类

为解决磁共振(magnetic resonance, MR)脑部图像来源不一以及病变位置和形态不固定造成MR脑部图像分类精度不高的问题,提出基于局部二值模式(local binary pattern, LBP)的纹理特征提取,并用极限学习机(extreme learning machine, ELM)对MR图像分类。计算图像感兴趣区域(region of interest, ROI)的掩码,将图像分成扇形的子区域,统计掩码坐标下各块子区域的LBP直方图,连接所有LBP直方图作为特征向量通过ELM进行分类。相比以前的方法,该方法能够计算颅脑内局部纹理特征,能分类来源不一以及多种病变的图像。对脑部MR图像分类进行试验,对所有样本分类正确率超过92%,正类样本正确率超过93%,负类样本正确率超过91%。试验结果表明,该方法能够对较为复杂的MR图像进行正确分类。

MR image classification based on LBP and extreme learning machine

To solve the problem that theMR brain images are collect from different sources and the pathological fields are varied, a method combining the texture feature extractor which was based on the local binary patterns(LBP)with the extreme learning machine(ELM)classifier was proposed. Mask for region of interest(ROI)was calculated, the image was divided into some sector subareas, LBP histograms were calculatedin every subarea, all the LBP histograms were connected as feature vector and then classified through ELM.Compared with previous methods, the new method could calculate local features, and it was feasible to classify the different sources of MR images and variously lesion images. Some experiments for MR image classification were done, and the accuracy was more than 92% for all samples, the accuracy was more than 93% for positive sample, the accuracy was more than 91% for negative sample. The results showed that the method was available for the varied MR images.