基于改进ReliefF与ACO特征选择算法的心肌病分类模型

Optimized Feature Selection Algorithm Based on ReliefF and Ant Colony for Cardiomyopathy Classification

基于MR图像来进行心肌病模式的识别以用于医生辅助诊断。首先提出了一种心肌图像重构的方法,该方法以心肌MR图像的医学结论为基础,采用线性插值的方式将原始心肌重构为更能表现环方向特征的形式,通过灰度共生矩阵(GLCM)、灰度游程矩阵(GLRLM)、局部二值模式(LBP)以及直方图模式提取出422个纹理特征。然后针对特征维数高的情况,提出了一种改进ReliefF算法和ACO算法的联合特征选择方法,引进欧氏距离为基础的距离系数改进ReliefF,所得的特征权重配合皮尔逊相关系数、识别精度以及特征子集长度来改进ACO算法的信息素更新和剪枝策略等。该算法在4个公开的高维特征基因数据集上,平均选择0.4%的特征,平均精度可达91.73%。该算法在重构和原生所构成的纹理特征中选取了6个特征用于三种心肌纤维化模式(正常、HCM、DCM)的识别,通过SVM分类模型,在测试集上取得了93.8%的准确率,可在临床应用中辅助医生进行心肌病诊断。

Recognition of cardiomyopathy patterns based on MR images is used to assist diagnosis. First, a method of myocardial image reconstruction is proposed. Based on the medical conclusions of myocardial MR images, linear interpolation is used to reconstruct the original myocardium into a form that can better express the characteristics of the ring direction. 422 texture features are extracted through gray level co-occurrence matrix (GLCM), gray run length matrix (GLRLM), local binary mode (LBP) and histogram mode. Then, in view of the high sample feature dimension, a joint feature selection method of improved ReliefF algorithm and ACO algorithm is proposed. The distance coefficient based on Euclidean distance is introduced to improve ReliefF, and the obtained feature weights are combined with Pearson correlation coefficient, recognition accuracy and feature subset length to improve the pheromone update and pruning strategy of the ACO algorithm. The algorithm selects 0.4% of the features on 4 publicly available high-dimensional feature gene data sets, with an average accuracy of 91.73%. In the end, the algorithm selected 6 features from the texture features composed of reconstruction and original to identify the three myocardial fibrosis modes (normal, HCM, DCM). Through the SVM classification model, it achieved 93.8% on the test set. The accuracy rate can assist doctors in the diagnosis of cardiomyopathy in clinical applications.