基于机器学习方法的超声M模式气胸图像的分类研究

Classification of ultrasound M-mode pneumothorax images based on machine learning methods

本研究基于图像提取特征结合机器学习方法,建立超声M模式图像分类模型,为气胸诊断提供参考。收集肺部滑动存在特征典型图像171幅,特征不典型图像283幅;肺部滑动消失特征典型图像1113幅,特征不典型图像111幅;肺点特征典型图像850幅,特征不典型图像285幅。通过提取灰度共生矩阵、灰度游程矩阵等纹理特征,采用五折交叉验证方法,使用随机森林、朴素贝叶斯和支持向量机3种分类器对M模式下超声图像进行分类。在使用支持向量机下,对单独特征典型图像进行分类的准确率最高,达到99.2%,灵敏度为99.54%,特异性为97.08%。实验结果证明,机器学习有望作为一种新的辅助诊断手段,有助于提高急救场合下的超声诊断气胸的准确率。

Based on image extraction features combined with machine learning methods,We established an ultrasound M-mode image classification model to provide a reference for pneumothorax diagnosis.We collected 171 images with typical features of lung sliding,283 images with atypical features;1113 images with typical features of lung sliding disappearing,111 images with atypical features;850 images with typical features of lung point,285 images with atypical features.We classified ultrasond images in M-mode by extracting texture features such as gray-level co-occurrence matrix and gray-level run-length matrix,using method of five-fold cross validation and three kinds of classifiers-random forest,naive bayes and support vector machine.Under the use of SVM,the classification of images of typical features had the highest accuracy of up to 99.2%,with a sensitivity of 99.54%and a specificity of 97.08%.The results of experiment show that machine learning is expected to be used as a new auxiliary diagnosis method to help improve the ultrasound diagnosis accuracy of pneumothorax in emergency.