| Radiomics is a hot topic in radiology in recent years.After high-throughput features extraction from medical images,such as CT,and artificial intelligence analysis,radiomics can provide valuable information for accurate diagnosis and evaluation of diseases.However,radiomics research still lacks perfect standards.The reproducibility of results of radiomics research is greatly challenged.Although some standards and recommendations for radiomics research have been proposed by some researchers,they still need to be improved.Wavelet features are calculated from wavelet decomposing images.They contain high-order information of images.They are important parts of radiomics features and they have been widely used in radiomics research.However,there is no standard and suggestion for the calculating wavelet features in radiomics,especially the selection of wavelet bases.The influence of different wavelet bases on the features calculated from decomposing images is not clear.Therefore,this study will study wavelet features from CT images of colorectal cancer to explore the effects of on the stability and diagnostic performance of wavelet features extracted form different wavelet filters.It could provide some suggestion on the use and selection of wavelet bases for radiomics research.The study collected preoperative CT images and clinical data of 256 patients with colorectal cancer,and divided the cases into training group(180 cases)and validation group(76 cases).After completing the delineation of the CT image tumor area,91 origin features and 52 wavelet feature sets from different wavelet base were extracted on an in-house MATLAB software.Based on the training group data,the stability of 616 features with the same name in different wavelet feature sets is analyzed.Stable featues may contain similar information of images.The stability of the feature is judged by calculating the intra-class correlation coefficient(ICC)of any two wavelet feature sets corresponding to the same-name feature.The feature,with ICC greater than 0.8,is judged as a stable feature.The number of stable features between the two feature sets while the times each feature is determined as a stable feature is recorded.The results show that the larger the difference between the wavelet orders,the fewer stable features between the corresponding feature sets,and less similar information between feature sets.The features of the non-uniformity metric and the features of the neighborhood gray tone difference matrix(NGTDM)are stable between different wavelet decomposing.In the training group,based on different wavelet feature sets,the least absolute shrinkage and selection operator(LASSO)was used to built radiomics model which respectively predict the statue of Ki-67,EGFR(epidermal growth factor receptor)gene mutation,and lymph node metastasis.The diagnostic performance of the model was verified in the validation group.The area under the receiver operating characteristic curve was used to evaluate the diagnostic performance of models.The results showed that wavelet features from some wavelet bases can improve performance of radiomics model.The models with best diagnostic performance for predicting different indicators are built form different wavelet feature sets,respectively.By analyzing the diagnostic performance of the models built from combined feature sets,it is found that too many features will lead to over-fitting and reduces the diagnostic performance.Based on the above analysis,this study respectively proposes a potentia diagnostic model for the three indicators respectively and an appropriate set including 7 wavelets.In this study,by analyzing of the stability and diagnostic performance of different wavelet feature set,the suggestions for calculating wavelet features in radiomics research are proposed.Wavelet bases with large differences in order can be selected together.One appropriate set includes 7 wavelets(db1,db5,db10,sym8,bior3.3,bior1.5 and rbio3.1).Features including similar information such as non-uniformity and NGTDM features should be avoided repetitive calculation to avoid feature redundancy and model over-fitting.In a radiomics study the diagnostic performance of each single wavelet feature set can be analyzed first and the five wavelet feature sets with the best diagnostic performance can be chosen and combined appropriately to obtain the model with the best diagnostic performance. |
PDF Full Text Request