Radiomics Study On Differential Diagnosis Of Colorectal Liver Metastasis And Its Correlation With Microsatellite Instability

PartⅠ: Research on machine learning-based radiomics features and model in identifying colorectal liver metastasis and breast cancer liver metastasisObjective Apply machine learning-based radiomics features and model to identify colorectal liver metastasis and breast cancer liver metastasis,and to evaluate the diagnostic efficiency of the model.Methods A total of 112 patients with CRLM and 118 patients with BCLM wereselected.And they were divided into training group andvalidation group according to the ratio of 7: 3 randomly(training: 160 cases,validation: 70 cases).The tumor’s region of interest in the enhanced CT(portal phase)image data of patientswas manually outlined.And 895 radiomics features were extracted.T-testand least absoluteshrinkage and selection operator(LASSO)were usedto perform feature dimension reductionrespectively.And optimal featureswere screened out,then the logistic regression(LR)model was created.The area under the receiver operating curve(AUC)value,accuracyvalue,sensitivityvalue,specificityvalueand F1＿score were used to assess the diagnostic efficacy of the model.Results In the training group,there were 78 patients with CRLM and 82 patients with BCLM.In the validation group,there were 34 patients with CRLM and 36 patients with BCLM.After feature dimension reduction,10 radiomics features with the largest differences were screened out,which are original＿shape＿Elongation,original＿shape＿Maximum3DDiameter,original＿shape＿Sphericity,original＿firstorder＿Skewness,waveletHLL＿glrlm＿Gray Level Non Uniformity Normalized,wavelet-HLH＿glrlm＿Run Entropy,wavelet-HHL＿glszm＿Zone Entropy,wavelet-HHH＿glrlm＿Long Run Emphasis,wavelet-HHH＿glszm＿Zone Entropy and wavelet-LLL＿glszm＿Size Zone Non Uniformity Normalized.The LR model was used to identify CRLM and BCLM with AUC value of 0.94,accuracyvalue of0.86,sensitivity value of 0.87,specificityvalue of 0.91,F1＿score of 0.90 in the training group.LR model in the validation group to identify CRLM and BCLM with AUC value of0.91,accuracyvalueof 0.89,sensitivityvalue of 0.83,specificityvalue of 0.89,F1＿score of0.89.Conclusion Based on the enhanced portal vein phase CT images of CRLM and BCLM,we can effectively identify the liver metastases from two different location using radiomics features,which provides a new idea for finding the primary focus of metastatic cancer with unknown primary tumor.Part Ⅱ : Research on machine learning-based radiomics features and model in identifying MSI-H and MSI-L in colorectal liver metastasisObjective Apply machine learning-based radiomics features and model to identify MSI-H and MSI-L in colorectal liver metastasis,and to evaluate the diagnostic efficiency of the model.Methods A total of 12 patients with MSI-H CRLM and 96 patients with MSI-L CRLM were selected.And they were randomly divided into training group and validation group according to the ratio of 7: 3(training: 75 cases,validation: 33 cases).Take the enhanced CT(portal phase)image data of patients to manually outline the tumor’s region of interest,and extract 788 radiomics features.T-test and LASSO were used to perform feature dimension reduction.And optimal featureswith the largest differences were screened out.Then the random forest(RF)modelwas created using the optimal features.The area under the receiver operating curve(AUC)value,accuracyvalue,sensitivityvalue,specificityvalueand F1_score were used to assess the model’s diagnostic efficacy.Results In the training group,there were 8 patients with MSI-H CRLM and 67 patients with MSI-L CRLM.In the validation group,there were 4 patients with MSI-H CRLM and 29 patients with MSI-L CRLM.After feature dimension reduction,7 radiomics features with the largest differences were screened out,which are original_shape_Elongation,wavelet-LLH_glcm_Idn,wavelet-LHL_firstorder_Skewness,wavelet-LHL_glrlm_Short Run Low Gray Level Emphasis,wavelet-HLH_glszm_Small Area Low Gray Level Emphasis,waveletHHH_glrlm_Run Length Non Uniformity Normalized,wavelet-LLL_glszm_Small Area High Gray Level Emphasis.The RF model was used to identify MSI-H and MSI-L in CRLM,and 5-fold cross-validation was adopted.The result shows AUC value of 0.88,accuracyvalue of 0.85,sensitivity value of 0.85,specificity value of 0.92,F1_score of 0.88 in the training group.RF model in the validation group to identify MSI-H and MSI-L with AUC value of 0.75,accuracyvalue of 0.74,sensitivityvalue of 0.81,specificityvalue of 0.85,F1_score of 0.78.Conclusion The prediction model based on enhanced CT images of portal phase can effectively identify MSI-H and MSI-L CRLM,which can provide effective auxiliary way for clinical immunotherapy in the case of unknown MSI,so as todelay the progression of metastasisand prolong the survival time.