基于SVM模型参数优化的多模态MRI图像肿瘤分割方法

摘要

Objective To develop a method for tumor segmentation on multi-modality magnetic resonance (MR) images based on parameter optimization of SVM model. Methods Each one of the 4 sub-classifiers was trained using the feature information in mono-modality MR images and applied to the corresponding modality images. The classification results differed due to different information in the selected support vectors of the mono-modality images. By modifying the weight values of the error data points, we chose the best weight values of the sub-classifier to obtain a weighed combination SVM classifier of multi-modalities for use in MR image segmentation. Result This tumor image segmentation method was validated on the MR images of brain tumors in 34 patients and resulted in an average classification accuracy of 90.59%. Compared with the 4 mono-modality classifiers, multi-modality RBF kernel SVM classifiers increased the overall accuracy by 5.76%-20.11%. Conclusion The proposed method combines multi-modality images with SVM classifiers to allow accurate tumor image segmentation from MR images with a high precision.%目的：提出一种基于混合核函数SVM模型参数优化的多模态MRI图像肿瘤分割方法。方法对多模态MRI图像中单一模态的特征信息，分别使用混合核函数SVM方法训练出4个子分类器，对相应模态进行分割。由于不同模态图像选择的支持向量各有侧重，分割结果存在差异。通过迭代修改分割错误数据点的权值，优化选择SVM模型子分类器权重系数，得到多模态加权组合的SVM分类器模型，并应用于多模态MRI图像分割。结果34例MRI脑肿瘤病人图像数据，获得了90.59%的分割精度，与单一模态分类器方法、多模态高斯核函数SVM方法相比，平均分割精度提高5.76%~20.11%。结论本文方法结合多模态图像和SVM的优势，提高肿瘤分割准确率，分割性能好。