Decomposition And Analysis Of Dynamic Enhancement Magnetic Resonance Imaging For Predicting Molecular Subtypes Of Breast Cancer

In recent years,the incidence of breast cancer has increased year by year,affecting the health of patients and even endangering their lives.Breast cancer was identified as four molecular subtypes by immunohistochemistry after biopsy in order to guide individualized treatment and prognosis analysis.In the process of exploring new non-invasive and efficient identification methods,some studies have shown that image features were potential molecular subtype markers.However,malignant tumors are highly heterogeneous.And heterogeneity is not only among different tumors,but also within one tumor which may hinders the accurate identification of molecular subtypes of breast cancer.In order to analyze the correlation between intratumoral blood flow heterogeneity and molecular subtypes,dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI)of breast was decomposed based on kinetic compartment model which consists of many types of tissues,and DCE-MRI signals are composed of the blood flow activity of each tissue.Because of the influence of tissue heterogeneity and the limitation of image resolution,the blood flow active regions presented in DCE-MRI include not only the independent regions of each tissue,but also the shared regions among tissues.i.es the independent region and shared region of each tissue constitute the complete region of blood flow activity of the tissue.In this study,DCE-MRI of breast tissue was decomposed into three tissues5 sub-components of blood flow activity.The complete region of blood flow activity of each tissue was called sub-component principal region,the independent one was called sub-component pure region,and the shared one was called sub-component mixing region.Compared with other DCE-MRI clustering partitioning methods,the result shows that the image features extracted from DCE-MRI sequences of sub-component principal region have the strongest correlation with molecular subtypes,and the random forest prediction model got the best result based on this region.The specific research contents of this paper include:(1)Association analysis between clinical information and molecular subtypes in breast cancer patients.Sort out statistical data of clinical patients,including age,type of cancer,maximum diameter of tumors and menstrual status.The patients were divided into four groups according to their molecular subtypes.The differelces of clinical information among groups Uere analyzed using χ2-Fisher and ANOVA methods.The analysis did not reveal any clinical priori information significantly related to molecular subtypes.(2)Decomposition for breast region of interest on DCE-MRI using convex analysis method and other clustering partitions methods.The tumors and stroma regions were segmented using man-machine interaction method according to the imaging characteristics of DCE-MRI and the sectional anatomical structure of breast.After that,according to the kinetic compartment model,the region of interest(ROI)on DCE-MRI was decomposed by convex analysis method to determine the sub-component principal region,pure region and aliasing region of each tissue.Another way,ROI was clustered by other methods on the basis of time-intensity curve,and its sub-regions were attained for comparative analysis.(3)Analysis of correlation between image features of breast tissue sub-components and molecular subtypes.The image features of tumor and stroma sub-components were extracted,including the statistical and texture features.Logistic regression was utilized to analyze the correlation between individual image feature and molecular subtypes.The correlation between the subcomponent and molecular subtypes was quantitatively evaluated and compared with other clustering regions according to the number of highly correlated image features in each region.The results showed that the sub-component principal region had strongest correlation with the molecular subtypes.(4)Molecular subtypes prediction with DCE-MRI sub-component image features of breast tissue.Based on the random forest algorithm,taking the image features of different sub-components of breast tissue as input,a variety of molecular subtypes prediction models were constructed,then we evaluated and compared the predictive effect among the models.The results showed that the sub-component principal region of tumor was best in prediction of molecular subtypes,that was significantly better than that of the whole region of undissolved tumors or sub-regions determined by other clustering partitioning method.Furthemiore,the prediction ability of molecular subtypes was significantly improved when the prediction models of sub-component principal regions of both tumor and stroma was fused.The results of comparative experiments showed that the imaging features of DCE-MRI sub-component principal region of heterogeneous breast tissue had the strongest correlation with molecular subtypes,and the random forest prediction model based on that had the best predictive effect.Overall,DCE-MRI decomposition analysis of breast tissue can provide more effective imaging information for identifying molecular subtypes in clinic,which is a potential molecular subtype imaging marker.