MG-based Radiomic Features Predict Lymph Node Metastasis Of Breast Cancer

Breast cancer is one of female malignant tumors,it has a very high probability to cause lymph node metastasis in breast cancer patients.And lymph node metastasis is one of the most important factors which influence how long the breast cancer patient can be alive.Doctors can judge whether it has lymph node metastasis in breast cancer patient by sentinel lymph node biopsy clinicaly.Biopsy is one kind of invasive test,and the result of sentinel lymph node biopsy can only be get at regular time.What's more,some patients are unsuited to take sentinel lymph node biopsy because of some unchangeable factors,for example,a patient who has other illness or the number of his sentinel lymph node is more than common people,or one patient who has to require additional operative time.An accurate and noninvasive detection method is needed to evaluate whether a patient has lymph node metastasis in pre-operation.During recent years,along with the development of machine learning technology a method called radiomics has been born.This method is to transform the area-of-interest of medical pictures into a high dimensional feature space and dig out potential pathological changes of tissues from medical images by analyzing these characteristic data.Some scholars have used this method to build colon cancer lymph node metastasis prediction model successfully.I suppose we can use radiomics method to analyze MG data and build a model to predict lymph node metastasis in breast cancer patients.The data contains 147 breast cancer patients whose Molybdenum target images were taken before surgery at Imaging Department in Shaanxi Provincial People's Hospital in the period 2016-2017.The methods are: Firstly,In accordance with the professional imaging physician's marking of the tumor region,a 3D-Slicer interactive software platform is used to segment the tumor region in the molybdenum target image of breast cancer patients,then transform the area-of-interest into quantization matrix,and extract features based on histogram and textural features calculated by Gray-level Co-occurrence Matrix and Gray Run-Length Matrixfrom private images and wavelet transformation-based images;Secondly,103 cases were randomly selected from the whole data included in this study to form the training set to construct the prognostic model,use 10 kinds of feature selection methods(Fisher score,Relief,Gini index Mutual information feature selection,Minimum redundancy maximum relevance,Conditional mutual information,T-score,Conditional mutual infor-mation maximization,Mutual information maximization,Double input symmetric relevance sort these features by relevance principle,take notes on the first 35 features in every feature selection method records,7 kinds of classification methods(Logistic Regression,Naive Bayes,Decision Tree,K Neighbors,Support Vector Machines,Gradient Boosting Decision Tree,Random Forest)were used in the training process of the prognostic model,meanwhile,the number of features in the training process was changed to find an optimal construction strategy for each selector-classifier combination;Thirdly,the remaining 44 cases were used as the test set to evaluate the performance of these prediction model,select the best model which has good prediction performance.After compare the prediction performance of all models,we find the number of feature,feature selection method,and classification algorithm are showed significant impact upon the prediction performance of the model.Finally,a prediction model which has good prediction performance have been choosed,this model is constructed of ten features(T stage,Sum Average,Cluster Tendency,Autocorrelation,et al),and its testing accuracy can receive 0.82,testing AUC can receive 0.8.This model has certain reference value for preoperative noninvasive assessment of lymph node metastasis risk in breast cancer patients.