The Value Of MRI Radiomics In The Differential Diagnosis Of Non-edematous Brain Metastases And Lacunar Infarction

Part I Differentiation of non-edematous brain metastases from lacunar infarction based on MRI imaging features and radiomicsObjective:To explore the value of MRI imaging features and MRI-based radiomics in the differential diagnosis of non-edematous brain metastases and lacunar infarction,in order to improve the understanding and diagnostic accuracy of these two diseases.Methods:The complete clinical and imaging data of 217 patients with non-edematous brain metastases and 206 patients with lacunar infarction confirmed by surgical pathology or clinical and imaging follow-up in the First Affiliated Hospital of Wannan Medical College from January 2015 to October 2022 were retrospectively analyzed.There were 131 males and 86 females in the brain metastasis group,aged 40-80(63.01±8.18)years;there were132 males and 74 females in the lacunar infarction group,aged 41-79(64.26±8.41)years.All patients underwent conventional head MRI plain scan,enhanced scan and DWI.The general clinical data and MRI imaging characteristics of all patients were observed and recorded,including gender,age,maximum diameter of tumor,shape(round-like/irregular shape),boundary(clear/fuzzy),T1WI signal ratio,T2WI signal ratio,T2-FLAIR signal ratio,DWI signal ratio and ADC value.The above two groups of cases were randomly divided into the training set(297 cases)and the test set(126 cases)according to 7:3.Two independent samples t test,Mann-Whitney U test andχ~2 test were used to compare the clinical and MRI features between the two groups.The feature variables with P<0.05 were included in the multivariate logistic regression analysis and the imaging feature model was established.ITK-SNAP software was used to manually delineate the lesions from axial MRI T2-FLAIR sequence images and fuse them into a three-dimensional region of interest(VOI).Then the original images and VOI images were imported into AK analysis software for radiomics feature extraction,and 1316 features were extracted for each lesion.The minimum redundancy maximum correlation(m RMR)and least absolute shrinkage and selection operator(LASSO)algorithms were used for dimensionality reduction and feature selection to obtain the most valuable radiomics features.The Rad-score of each patient was calculated according to the linear fitting equation and the radiomics model was constructed.The statistically significant imaging features were combined with Rad-score to construct a joint model and its nomogram.The receiver operating characteristic(ROC)curve and the area under the curve(AUC)were used to evaluate the differential diagnostic efficacy of each model,and the test set data were used for internal validation.The Delong test was used to evaluate the differences in efficacy between the models.Calibration curves were used to evaluate the agreement between the predicted values of the model and the observed values.Decision curve analysis(DCA)was used to determine the clinical application value of each model.Results:There were significant differences in DWI signal ratio and the lesion margin between the two groups(P<0.05).The statistical distribution of DWI signal ratio in brain metastasis group was more concentrated than that in lacunar infarction group.The edge of the brain metastasis group was more clear,while the edge of the lacunar infarction group was more fuzzy.The AUC,sensitivity and specificity of the model based on lesion edge features and DWI signal ratio in the training set and test set were 0.690,93.4%,41.4%and0.710,90.8%,42.6%,respectively.A total of 1316 radiomics features were extracted from each lesion.After m RMR and LASSO dimensionality reduction screening,15 most valuable radiomics features were retained,including 6 first-order features and 9 texture features,among which the original＿firstorder＿Kurtosis feature had the highest weight coefficient.The AUC value,sensitivity and specificity of the ROC curve of the radiomics model based on the 15 features were 0.829,59.9%and 92.4%in the training set and 0.812,83.1%and63.9%in the test set,respectively.The AUC values of the combined model based on imaging features and radiomics features in the training set and test set were 0.873 and 0.855,respectively,and the corresponding sensitivity and specificity were 85.5%,71.0%and81.5%,73.8%,respectively.Delong test results showed that the differences in AUC values among the models were statistically significant.The calibration curve showed good agreement between the predicted values and the observed values of the combined model.DCA curve showed that the combined model had the highest net clinical benefit.Conclusion:1.The DWI signal ratio and lesion margin of MRI imaging features have certain value in the differential diagnosis of non-edematous brain metastases and lacunar cerebral infarction,but the imaging feature model based on the two features has poor performance;2.Radiomics based on MRI can differentiate non-edematous brain metastases from lacunar cerebral infarction,which is difficult to be differentiated by conventional MRI;3.Among the three models,the diagnostic efficacy(AUC)of the combined model was0.873 in the training set and 0.855 in the test set,which was better than the individual imaging feature model and radiomics model,and had good clinical application value.Part II Different machine learning models based on MRI T2-FLAIR sequence radiomics in differentiating nonedematous lung adenocarcinoma brain metastases and lacunar infarctionObjective: To analyze and compare the value of radiomics models established by different machine learning classification algorithms based on MRI T2-FLAIR sequence in the differential diagnosis of brain metastases from non-edematous lung adenocarcinoma and lacunar infarction.Methods: The pre-treatment magnetic resonance images of 104 patients with brain metastases from lung adenocarcinoma and 165 patients with lacunar infarction confirmed by surgical pathology or clinical and imaging follow-up in the First Affiliated Hospital of Wannan Medical College from January 2015 to October 2022 were retrospectively analyzed,according to the signal level of DWI sequence,the patients were divided into two groups(105 cases in the hyperintensity group and 164 cases in the iso-low signal group).All patients underwent conventional head MRI plain scan,enhanced scan and DWI.Each group was randomly divided into training set and test set at a ratio of 7∶3.Radiomics features were extracted from manually delineated 3D regions of interest on axial T2-FLAIR sequence images.Using m RMR and LASSO regression for dimensionality reduction,the radiomic features with the most diagnostic value were screened,and the models were constructed separately by combining four machine learning classifiers,and ROC curves were drawn,the diagnostic performance of each model was evaluated using the area under the curve(AUC),sensitivity and specificity.Results: The feature with the highest weight coefficient in the iso-low signal group is the First Order_Total Energy;the feature with the highest weight coefficient in the high-signal group is the Gray Level Size Zone Matrix_Small Area Low Gray Level Emphasis.Among the four machine learning classification models,the random forest(RF)model performed best in the iso-low signal group,the AUC,sensitivity and specificity were 0.887,0.892,0.772(training set)and 0.901,0.800,0.939(test set),respectively;the decision tree(DT)model performed best in the high-signal group,whose AUC,sensitivity and specificity were 0.838,0.892,0.605(training set)and 0.816,0.800,0.733(test set),respectively.Conclusion: 1.The radiomics models established by different machine learning classification algorithms based on T2-FLAIR sequence have certain value in differentiating non-edematous lung adenocarcinoma brain metastases and lacunar infarction.2.The differential diagnostic efficacy of each classification model was different.The best in the iso-low signal group was the random forest model,and the best in the high signal group was the decision tree model.