基于数字胸片小波纹理特征的尘肺病早期诊断方法研究

目的探讨利用基于小波变换的熵纹理特征进行尘肺病诊断的方法，并研究相关的分类技术。方法对70名健康体检者和40名尘肺病患者的数字x射线摄影（digitalradiography，DR）图像进行纹理分析，提取小波熵纹理特征，并利用决策树进行特征选择。选取不同核函数的支持向量机（supportvectormachines，SVM）对DR胸片进行分类，通过5折交叉验证估计诊断分类的性能并进行评价。结果对DR图像做8次小波分解后提取8个小波熵纹理特征（特征全集），其中6个经过特征选择组成特征子集。应用SVM进行分类时，基于特征子集的分类结果均好于基于特征全集的分类结果。线性核函数SVM的分类效果好于其他核函数SVM的分类效果，准确率达84．6％，ROC曲线下面积为0．88±0．04。结论利用SVM以DR图像的小波熵为特征进行尘肺病诊断有较高水平，有助于尘肺病的早期诊断。

Early diagnosis of pneumoconiosis on digital radiographs based on wavelet transform-derived texture features

Objective To investigate the early diagnosis of pneumoconiosis on digital radiographs by means of wavelet transform-derived entropy and the related technologies of classification. Methods Wavelet transform-derived entropies were extracted from the digital X-ray radiographies（DRs） of 70 normal persons and 40 pneumoconiosis patients and were selected by decision tree. Support vector machines（SVMs） with different kernel functions were adopted to distinguish pneumoconiosis DRs from normal DRs. The classification performance was estimated and evaluated through 5-fold cross validation. Results The DR images were wavelet- discomposed for 8 times,resulting in 8 wavelet entropies to form the feature full-set, and six were selected to form the feature subset. The classification performances based on the feature subset were better than those based on the feature full-set when classification was done with SVMs. SVM with linear kernel function performed better than SVMs with polynomial and Gauss kernel functions,with accuracy of 84.6% and an area under the ROCcurve of 0. 88±0.04. Conclusions The early diagnosis of pneumoconiosis based on wavelet transform-derived texture features with SVM is of a high level.