Analysis Of Tumor Heterogeneity Based On DCE-MRI Deconvolution Model And Its Application In Predicting Molecular Subtypes Of Breast Cancer

Breast cancer has high incidence rate and high mortality rate,which seriously endangers the life and health of women.Breast cancer is divided into four molecular subtypes by immunohistochemical method.There are great differences in clinical manifestations of breast cancer with different molecular types.Molecular typing is expected to be the basis for formulating individualized treatment plan in the future.However,the high heterogeneity of breast cancer affects the accurate molecular typing.Dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI)can not only obtain the morphological information of the lesions,but also reflect the heterogeneous microscopic information such as tissue vascular permeability and local regional blood perfusion,which can be used in molecular typing prediction.Based on the atrioventricular dynamics model,the DCE-MRI image signal was abstracted as a weighted mixture of multiple atrioventricular pharmacokinetic signals.The potential signal patterns were decomposed from the DCE-MRI images of breast cancer patients by cam deconvolution model.According to the differences of signal patterns of different patients,the molecular classification prediction was studied.The specific research contents of this paper include:(1)Processing and analysis of clinical pathological information and image data.The pathology of 205 patients with breast cancer was retrospectively analyzed.The clinical case information and DCE-MRI image data were collected.The correlation between the basic clinical information including patient’s age,menopause,tumor type and tumor maximum diameter was analyzed by ANOVA,chi square test and Fisher’s exact test.The tumor and gland were segmented as regions of interest by semi-automatic computer method,and preprocessed for subsequent image decomposition.(2)The research of tumor signal mode decomposition based on CAM deconvolution model.In order to explore the heterogeneity of blood flow in breast cancer and improve the prediction performance of breast cancer molecular classification,the CAM deconvolution model combined with bicubic interpolation was used to explore the blood flow heterogeneity of breast cancer and improve the prediction performance of breast cancer molecular typing DCE-MRI image signal decomposition was performed on the whole matrix of breast cancer patients,including the surrounding glands,to find the hidden signal patterns.The minimum description length(MDL)criterion was used as the evaluation index to evaluate the number of optimal signal modes.The preliminary correlation between signal patterns and molecular typing was studied.Finally,non-negative constrained least squares(NNLS)method is used to obtain the probability matrix of signal mode corresponding to each case in the unified signal mode,and the image is restored.(3)Based on the image decomposition features of CAM deconvolution model,the molecular typing association was studied.The probability matrix obtained by decomposition is the quantitative expression of different signal modes,and contains tumor heterogeneity information.Based on the principle of gray level co-occurrence matrix(GLCM),a signal mode co-occurrence matrix algorithm based on image decomposition is constructed and implemented,and the statistical features and texture features are extracted by probability matrix.The molecular characteristics of breast cancer were evaluated by molecular model and unimodal analysis.In single feature analysis,support vector machine(SVM)model is used to evaluate the extracted features one by one,and the correlation between single features and molecular typing of different signal modes is explored by statistical method.Multi-feature analysis uses Pearson dimension reduction to remove redundancy,uses support vector machine recursive feature elimination(SVM-RFE)to rank the importance of features,grid search method combined with ten-fold cross validation to determine the optimal super parameters of the model,select the optimal feature subset,and establish the multi feature prediction model based on the number of different signal modes.The prediction performance of the model was evaluated by AUC,and the 95% confidence interval of AUC was calculated by bootstrap method.(4)Molecular typing prediction of breast cancer based on ensemble learning region model feature fusion.In order to fully explore the interaction between tumor and surrounding glands,ensemble learning methods including average method,weighted method,voting method and stacking learning method were used to predict molecular typing combined with tumor,gland and tumor gland model.The experimental results show that the optimal AUC of Luminal A,Luminal B,Basal-like and HER-2 are 0.708,0.651,0.646 and 0.768 respectively when k = 4,and the optimal AUC of molecular typing obtained by probability weighting method for tumor,parenchyma,tumor &parenchyma models are 0.774,0.701,0.658 and 0.804 respectively.The results show that the potential signal pattern features can be decomposed from DCE-MRI images of breast cancer patients by cam deconvolution model,which has certain prediction performance of molecular classification.The combination of tumor,parenchyma,tumor & parenchyma multi region model can improve the prediction performance of molecular classification of breast cancer.