不同类型小波滤波对影像组学特征相关性和诊断效能的影响

Effects of different wavelet filters on correlation and diagnostic performance of radiomics features

目的:探讨不同小波滤波对影像组学特征相关性和诊断效能的影响。方法:回顾性收集143例结直肠癌患者(淋巴结转移阳性64例,阴性79例)的术前CT图像,经放射科医师勾画肿瘤区域后,使用Matlab编写的软件提取不同类型小波的影像组学特征。通过计算相关系数分析不同小波间同名特征的相关性。采用最小绝对收缩和选择算子(the least absolute shrinkage and selection operator,LASSO)构建不同的小波特征集预测淋巴结转移的影像组学标签并采用Delong’s检验比较其效能。结果:随着小波阶数差异的增大,小波间高相关同名特征数量减少。部分特征在不同小波间易出现高相关性。单个小波的特征集中rbio2.2,sym7和db7的特征子集构建的影像组学标签诊断效能最高。Daubechies系列小波特征集构建的标签预测淋巴结转移效能最高,Biorthogonal系列小波标签则最低,在去除同名高相关特征后全体特征集的诊断效能显著提高(P=0.004)。结论:建议选择阶数差异大的小波以降低影像组学特征的数据冗余度。为提高标签的诊断效能,有必要去除高相关特征。

Objective:To investigate the effects of different wavelet filters on correlation and diagnostic performance of radiomics features.Methods:A total of 143 colorectal cancer(CRC)patients(64 positive in lymph node metastasis and 79 negative)with contrast-enhanced CT examination were recruited.After labeling the tumor area by experienced radiologists,radiomics wavelets features based on 48 different wavelets were extracted using in-house software coded by Matlab.The correlation coefficients of the features withsame names between different wavelets were calculated and got the distribution of high-correlation features between each wavelet.The least absolute shrinkage and selection operator(LASSO)was used to build signatures between lymph node metastasis and wavelet features data set based on different wavelets.The numbers of features in signatures and diagnostic performance were compared using Delong’s test.Results:With the difference of wavelet order increased,the number of high-correlation features between two wavelets decreased.Some features were prone to high correlation between different wavelets.When building radiomics signature based on single wavelet,signatures built from‘rbio2.2’,‘sym7’and‘db7’did well in predicting lymph node metastasis.The signature based on Daubechies wavelet feature set had the highest performance in predicting lymph node metastasis,while the signature from Biorthogonal wavelet features was worst.Improvement was significant in diagnostic performance after excluding the high-correlation features in the whole features set(P=0.004).Conclusion:In order to reduce the data redundancy of features,it is recommended to select wavelets with large differences in wavelet orders when calculating radiomics wavelet features.It is necessary to remove high correlation features for improving the diagnostic performance of radiomics signature.