The Application Of Proton Exchange Rate Magnetic Resonance Imaging And Multimodal Radiomics In Predicting The Outcome Of Ischemic Stroke

Objective:Ischemic stroke（IS）is a serious cerebrovascular disorder worldwide with high morbidity and disability,approximately accounting for 80%of all strokes.Early prediction of the functional outcome of IS is of great importance to facilitate individualized treatment.This study aimed to evaluate the predictive value of the proton exchange rate(k_ex)magnetic resonance imaging（MRI）for the functional outcome of acute ischemic stroke（AIS）and construct a clinical-radiomics nomogram for predicting the functional outcome of AIS patients to assist in clinical decision-making.Automated machine learning and multimodal radiomics were used to further enhance the predictive performance.Methods:A total of 96 IS patients and 30 healthy subjects were recruited to undergo k_exMRI in this study.k_ex maps were calculated by water direct saturation-removed omega plots.The ratio of k_ex in the stroke lesion to the contralateral mirror tissue was calculated to generate the relative k_ex(rk_ex).The predictive performance of k_ex and rk_ex on the outcome of AIS were evaluated using the receiver operating characteristic curve（ROC）.In addition,522 patients with AIS were randomly divided into the training（n=311）and testing cohorts（n=211）.According to the modified Rankin Scale（m RS）at six months after hospital discharge,their functional outcome was dichotomized into good（m RS≤2）and poor（m RS>2）.Radiomics features were extracted from diffusion-weighted imaging（DWI）and apparent diffusion coefficient（ADC）maps.The radiomics model,clinical model and clinical-radiomics nomogram were constructed through feature selection and multivariate logistic regression.The performance of the models was evaluated and compared using calibration curve,ROC,decision analysis curves and Delong test.Another 203 patients with AIS were randomly divided into the training and testing cohorts at a ratio of 8:2.Their functional outcome was assessed based on the m RS score.Radiomics features were extracted from diffusion kurtosis imaging（DKI）,DWI,ADC maps,and T2 fluid-attenuated inversion recovery,respectively.The outcome prediction models based on different feature sets were developed using automated machine learning.The performance of models was assessed using ROC,accuracy,and decision analysis curves.Results:The k_ex of the stroke lesions was significantly higher than that of the contralateral mirror tissue and white matter of healthy subjects.The k_ex of lesions was significantly higher in AIS patients with poor prognosis than in those with good prognosis.The area under the curve（AUC）of k_ex and rk_ex for prediction of AIS functional outcome was 0.837 and 0.880.Age,sex,stroke history,diabetes,baseline m RS,baseline National Institutes of Health Stroke Scale（NIHSS）score and radiomics score were independent predictors of AIS functional outcome.The predictive performance of the clinical-radiomics nomogram was significantly higher than that of the simple clinical and radiomics model,with AUC of 0.868and 0.890 in the training and testing cohorts,respectively.The nomogram fitted well in the calibration curves（P>0.05）.In addition,eight different prediction models were developed by automated machine learning based on different combinations of feature sets.The DKI-based radiomics-clinical model yielded the greatest predictive performance in the testing cohort（AUC=0.923,average precision=0.896,accuracy=0.775,sensitivity=0.800,specificity=0.760）.Conclusions:The k_ex of the stroke lesions was related to the prognosis of AIS patients.The clinical-radiomics nomogram developed in this study could assist clinicians in predicting the risk of poor functional outcome of AIS patients,so as to formulate individualized treatment plans for different patients and improve the ultimate outcome.Automated machine learning based on multimodal radiomics could further improve the predictive performance of models,and is of great significance to promote accurate prediction of AIS outcome.Part Ⅰ: The Application of the Proton Exchange Rate Magnetic Resonance Imaging（kex MRI）in Ischemic StrokeObjective: kex MRI has recently been developed and preliminarily shown its potential value for evaluating reactive oxygen species.This study aimed to investigate the kex in different stroke stages and its correlation with stroke severity and prognosis.Methods: 96 IS patients were included in this study and divided into 3 groups based on stroke phases（acute,subacute and chronic）.In addition,30 healthy subjects were recruited in this study.A spin-echo echo-planar imaging sequence with saturation radio frequency power of 1.5,2.5,and 3.5 μT was implemented to obtain Z-spectra.kex maps were constructed from water direct saturation-removed omega plots.rkex and relative apparent diffusion coefficient（r ADC）were calculated by taking the ratio of kex or ADC in the infarcts over that of the contralateral tissue,respectively.Pearson correlation analysis was implemented to evaluate the correlations between both kex and rkex and NIHSS score.kex,rkex,r ADC,and lesion volume were evaluated for predictive performance of acute stroke outcome using the ROC.Results: The kex of ischemic lesions was significantly higher than the contralateral mirror tissue and white matter of healthy subjects.The difference between the kex of contralateral brain tissue of IS patients and the white matter of healthy subjects was also statistically significant.Besides,the kex of acute lesions was higher than subacute and chronic lesions（935.08 ± 81.45 s-1 vs 881.42 ± 55.66 s-1,P < 0.05;935.08 ± 81.45 s-1 vs 866.93 ± 76.66 s-1,P < 0.01）.However,the difference between subacute and chronic lesions was not significant.The kex and rkex of acute lesions showed positive correlation with the NIHSS score（r = 0.406,P = 0.016;r = 0.531,P = 0.001）.Acute patients with poor prognosis had significantly higher kex and rkex of lesions compared to patients with good prognosis（991.08 ± 78.22 s-1 vs.893.08 ± 55.06 s-1,P < 0.001;1.28 ± 0.09 vs.1.15 ± 0.06,P < 0.001）.These measures showed favorable predictive performance for AIS outcome with AUC of 0.837 and 0.880,slightly while not significantly higher than lesion volume（AUC: 0.730）and r ADC（AUC: 0.673）.Conclusions: kex MRI is promising for diagnosis and management of AIS as it could reflect the severity and the risk of poor prognosis of AIS patients.Part Ⅱ: Construction of the Nomogram Combining Diffusion-Weighted Imaging-Based Radiomics and Clinical Factors for Predicting the Functional Outcome of Patients with Acute Ischemic StrokeObjective: To develop a nomogram incorporating DWI-based radiomics features and clinical factors for accurate prediction of the functional outcome of AIS patients.Materials and Methods: Data from 522 AIS patients which were retrospectively collected were randomly divided into the training（n = 311）and testing cohorts（n = 211）.According to the m RS at six months after hospital discharge,their functional outcome was dichotomized into good（m RS ≤ 2）and poor（m RS > 2）.Radiomics features were extracted from DWI and ADC maps.The minimum redundancy maximum relevance algorithm and the least absolute shrinkage and selection operator logistic regression method were implemented to select features and establish radiomics model.Univariate and multivariate logistic regression were performed to sift clinical factors and construct clinical model.Ultimately,a multivariate logistic regression analysis incorporating independent clinical factors and radiomics score was implemented to establish the clinical-radiomics model using backward stepdown selection procedure and a nomogram was developed.Its performance was evaluated by calibration curve,ROC,precision-recall（PR）curve and decision curve analysis.Delong test was used to compare the predictive performance of these three models.Results: Age,sex,stroke history,diabetes,baseline mRS,baseline NIHSS score and radiomics score were independent predictors of AIS functional outcome.The clinicalradiomics model yielded the greatest predictive performance with an AUC-ROC of 0.868 in the training cohort and 0.890 in the testing cohort,the AUC-PR reached 0.733 and 0.787 respectively.The nomogram fitted well in the calibration curves（P > 0.05）.Decision curve analysis indicated its clinical usefulness.Delong test demonstrated that the difference between AUC-ROC of the clinical-radiomics nomogram and the simple clinical or radiomics model is significant.Conclusion: The clinical-radiomics nomogram could accurately predict the functional outcome of AIS and assist clinicians to formulate individual treatment plans at the early stage of onset,which may improve the ultimate outcome of the AIS patients.Part Ⅲ: Automated Machine Learning Based on Clinical Factors and Multimodal Radiomics of Diffusion Kurtosis Imaging and Conventional Magnetic Resonance Imaging Sequences Predicts the Functional Outcome of Acute Ischemic StrokeObjectives: This study aimed to investigate the predictive value of the radiomics from DKI for the functional outcome of AIS patients and develop a prediction model based on multimodal radiomics and clinical factors using automated machine learning to accurately predict the functional outcome.Methods: The clinical and imaging data of 203 patients with AIS were retrospectively collected and randomly divided into the training cohort（n = 163）and testing cohort（n = 40）.In light of the m RS at three months after hospital discharge,the functional outcome was dichotomized into good（m RS ≤ 2）and poor（m RS > 2）.Radiomics features were extracted from DKI parametric maps and conventional MRI sequences including T2-weighted fluid-attenuated inversion recovery,DWI and ADC maps.The tree-based pipeline optimization tool（TPOT）was applied to establish the prediction models based on different feature sets.Then the best models were validated on the testing cohort.The AUC,average precision,accuracy,sensitivity and specificity were used to evaluate the predictive performance of the best models.Results: Four radiomics models and four radiomics-clinical models were developed by TPOT.The AUC of DKI-based radiomics model was higher than the models based on conventional MRI sequences.The multimodal radiomics models performed best among the radiomics models with the AUC of 0.888.The models combining radiomics score from DKI parametric maps and clinical factors achieved the best performance in the testing cohort（AUC = 0.923,average precision = 0.896,accuracy = 0.775,sensitivity = 0.800,specificity = 0.760）.Conclusions: The DKI-based radiomics could further improve the performance for predicting the AIS outcome compared with conventional MRI sequences.The models developed by TPOT based on multimodal MRI radiomics and clinical factors could predict the functional outcome of AIS patients with a high discriminatory accuracy.