Research On Predicting Distant Metastasis Of Breast Cancer Based On Magnetic Resonance Imaging Radiomics Features

Objective To establish a predictive model to predict distant metastasis of breast cancer based on the dynamic contrast enhanced magnetic resonance imaging.Methods We retrospectively reviewed the breast MRI of patients confirmed as breast invasive cancer between January 2011 and December 2016 in our hospital.A total of 93 cases who had distant metastases were investigated and the relationship between breast cancer subtypes and distant metastasis was analyzed.Then 186 patients who had not distant metastases during the follow up time were randomly selected in chronological order as a control group.93 cases with 97 lesions(3 cases of synchronous bilateral breast cancer and a case of multifocal breast cancer)with distant metastasis occurred as metastasis group,186 cases with 191 lesions(4 cases of multifocal breast cancer)without distant metastasis occurred as metastasis-free group.Breast tumors were volumetrically segmented from two groups based on the first phase of contrast-enhanced T1-weighted images,then the radiomics features from ROIs were extracted.The radiomics features included morphology,grayscale statistics,texture and wavelet features.Rank sum test and LASSO algorithm were conducted for feature dimensionality reduction,and the ensemble learning classifier was used to establish the predictive model.Further,the clinicopathological characteristics from two groups patients were collected.Statistically significant features were screened out by univariate analysis,and combined with radiomics features to establish a predictive model.The predictive effectiveness of the two models were compared by sensitivity,specificity,accuracy and AUC.Results There are 14 cases with Luminal A types(14.6%),50 with Luminal B types(52.1%),14 with Her-2 overexpression types(14.6%),8 with triple-negative types(18.7%)in the metastasis group.Because the immunohistochemical were incomplete,1 case without molecular Subtype.There are 23 cases with Luminal A types(18.7%),133 with Luminal B types(69.6%),11 with Her-2 overexpression types(5.8%),24 with triple-negative types(12.6%)in the metastasis-free group.The difference between the two groups was statistically significant(P<0.05).The most common distant metastatic sites for the 4 types of breast cancer were: Luminal A types had 8 cases of bone metastasis(57.1%),Luminal B types had 36 cases of bone metastasis(72.0%),Her-2 overexpression types had 9 cases of liver metastasis(64.3%),triple-negative types had 9 cases of lung metastasis(50.0%).We extracted the radiomics features including morphological,gray statistical features,texture and wavelet of the 288 breast cancer lesions.After carrying out the rank sum test and LASSO algorithm,we screened out 4 morphological features(Volume,sphericity,macrosphericity,concavity ratio),3 gray statistical features(gray span,mean,Fifth Central state),10 texture features(contrast,autocorrelation,entropy,angular second moment,inertia,inhomogeneity of gradient distribution,energy,gradient mean square deviation,contrast,angular second moment),and 16 wavelets.Univariate analysis showed that history of lactation,abortion,birth,number of births,menstruation,lymph node metastasis,ER,PR,Her-2,molecular subtypes were significantly differenet between these two groups(P≤0.05).The sensitivity,specificity,accuracy,and AUC of the model based on single radiomics features were 66.8%,84.3%,78.4%,and 0.755.When combining with clinicopathological characteristics the sensitivity,specificity,accuracy,and AUC of the model were improved to 67.8%,86.9%,80.6%,and 0.774.Conclusions Bone,lung and liver are the most common distant metastatic sites for breast cancer.Luminal A and Luminal B tumors have been shown to have a greater probability to develop bone metastasis.Triple-negative tumors show a greater probability to develop lung metastasis.Her-2 overexpression cases were more likely to develop liver metastasis.The two predictive models established in this study can effectively predict the distant metastasis of breast cancer,and combining radiomics features with clinicopathological features have better predictive performance.The two predictive models have potential value in the development of clinical treatment options.