Preliminary Clinical Application Of MRI Radiomics In The Diagnosis Of CO Poisoning Encephalopathy And Nomogram Prediction Of DEACMP

Objective:To explore the clinical application value of MRI radiomics model in the diagnosis of CO poisoning encephalopathy and clinical radiomics nomogram in predicting delayed encephalopathy after carbon monoxide poisoning（DEACMP）.Methods:This study prospectively analyzed the clinical and imaging data of 90 patients with CO poisoning diagnosed in the First Hospital of Lanzhou University from May 2018 to May 2022.All patients with CO poisoning underwent conventional MRI scans and 3D-T1WI scans.Among them,52 patients with acute CO poisoning were followed up for 3 months.The occurrence of DEACMP was recorded,and the clinical and imaging data of DEACMP group and non-DEACMP group were analyzed.For each participant,brain tissue was automatically segmented based on 3D-T₁WI imaging on Intelli Space Discovery（ISD）.The whole white matter,whole gray matter,corpus callosum,bilateral globus pallidus,bilateral caudate+nucleus accumbens,bilateral thalamus were selected as ROIs.The radiomics features of each ROI were automatically extracted using the Pyradiomics package.Then,two-sample independent t-test,Lasso regression and Pearson correlation matrix were used for feature selection,and the most valuable omics features were selected to construct radiomics diagnosis model and clinical radiomics nomogram.Each ROI used 9 classifiers to train the radiomics diagnostic model.The AUC values of the radiomics diagnostic model based on each ROI were compared,and the model with the largest AUC value was selected as the ideal MRI marker for diagnosing cerebral damage after CO poisoning.Univariate and multivariate logistic regression analysis was used to determine the independent risk factors that can predict DEACMP in clinical data and imaging features,and then the independent risk factors were combined to draw a clinical radiomics nomogram for predicting DEACMP.Finally,ROC curve,calibration curve and DCA curve were used to evaluate the accuracy and clinical application value of the nomogram.Results:1.Among the 90 patients with CO poisoning,40 patients had brain tissue lesions of CO poisoning on conventional MRI,and 50 patients had no obvious abnormalities.2.Follow-up of 52 patients with acute CO poisoning showed that 20 patients had DEACMP,of which 11 patients had brain lesions of CO poisoning on conventional MRI.Non-DEACMP occurred in 32 patients,of which 2 patients had brain tissue lesions on conventional MRI.3.A total of 1473 radiomics features were extracted from each ROI of 90 patients with CO poisoning.After feature screening,1-6 radiomics features were retained in each ROI for training the radiomics diagnosis model.4.Among the nine classifiers used,the support vector machine model has the best diagnostic performance.Whole gray matter had the best diagnostic efficiency,with AUC value,sensitivity and specificity in the validation group（0.815,0.888,0.714,respectively）and the test group（0.827,0.722,0.857,respectively）,followed by the left globus pallidus with an AUC value,sensitivity and specificity in the validation group（0.806,0.888,0.714,respectively）and the test group（0.806,0.777,0.714,respectively）.5.Univariate logistic regression analysis showed that there were significant differences in age,GCS score,MMSE score and brain tissue lesions between the two groups（P<0.05）.Multivariate Logistic regression analysis showed that MMSE score,GLSZM of wavelet transform,GLSZM of exponential transform,Skewness of wavelet transform and GLDM of Log transform were independent risk factors for DEACMP.6.The clinical radiomics nomogram AUC was 0.875.The calibration curve shows that the DEACMP predicted by clinical radiomics nomogram is consistent with the actual DEACMP;the DCA curve shows that the clinical net income of nomogram is poor.Conclusion:The MRI radiomics model based on whole gray matter and left globus pallidus can diagnose cerebral damage of CO poisoning.The clinical radiomics nomogram based on MMSE score,GLSZM,Skewness and GLDM can predict the occurrence of DEACMP.