Computational protocol: DWI Intensity Values Predict FLAIR Lesions in Acute Ischemic Stroke

Similar protocols

Protocol publication

[…] Co-registration and post-processing of DW- and FLAIR images was performed with VINCI, Version 2.63 (Max-Planck-Institute for Neurological Research, Cologne, Germany) . DWI lesion volumina were assessed using MRIcron (Chris Rorden, http://www.mccauslandcenter.sc.edu/mricro/). At 3 T, DW images were resized in the z-axis to match FLAIR images.In DWI, the slice with the largest lesion extent was identified visually and used for the complete further analysis. DW-images were co-registered to FLAIR images and the absence or presence of FLAIR lesions was assessed by three raters blinded to clinical data and DW-images (experience in stroke imaging is indicated for each rater; Rater 1, VIM: 3 years, Rater 2, ME: 5 years, Rater 3, JS: 10 years). Prior to the rating, raters were encouraged to look for subtle intensity changes by adjusting contrast and brightness of the images and to compare intensities of potentially hyperintense regions with the healthy contralateral hemisphere. Such subtle intensity changes were also rated as a FLAIR hyperintensity. The area of the FLAIR lesion was individually delineated by each rater and copied on the DW-images. Then, 6 mm regions of interest (ROIs) were placed within the whole DWI-lesion. Each ROI was labeled according to its position in regard to the FLAIR-ROI. If it was located inside the FLAIR lesion, it was labeled FLAIR+ (positive), if it was located outside of the FLAIR lesion, it was labeled as FLAIR- (negative). If a FLAIR lesion was absent, all DWI ROIs of this patient were labeled as FLAIR-. For a graphical overview of the analysis see . ROI-values were normalized as a ratio: [100% x (mean ROI value/mean value of the unaffected hemisphere)], taken from a slice at the height of the lateral ventricles and above the putamen encompassing the corona radiata. In cases of cerebellar infarction, a ROI of the contralateral cerebellar hemisphere was used to normalize the ROI values. Above steps were performed equally for apparent diffusion coefficient (ADC)-maps. [...] Owing to skewed distribution of some variables, results are presented as median and interquartile range (IQR) if not indicated otherwise. Differences in clinical data between groups were assessed using the Mann-Whitney U rank sum test.Agreement between raters for the identification of FLAIR-hyperintensities was analyzed using free-marginal kappa , . Kappa values were evaluated as suggested by Landis and Koch .ROI-analysis was performed on patient level and separately for the two centres. We used mean ROI-values per patient and per rater for positive ROIs and negative ROIs separately:a) If all raters had some ROIs of a patient classified as having a positive FLAIR status, we used only the mean of the positive values and classified the patient as having a positive flair status.b) If not all raters found positive ROIs for a patient, but all raters had ROIs classified as negative, we used only the mean of the negative ROI-values and classified the patient as negative. *c) Only for some patients one of the raters classified all ROIs in another category as the other raters. For them we used the classification of the two corresponding raters and set the ROI value for the rater not corresponding to missing.In the next step, the ability of a relative DWI-intensity threshold to predict the presence of corresponding FLAIR-hyperintensities was analyzed using an unadjusted receiver operating characteristics (ROC) curve analysis. The area under the curve (AUC) and the 95% confidence limits for the raters are reported. To get the optimal threshold, the Youden Index was used ( . Sensitivity, specificity, and predictive values for the optimal thresholds are also reported.To adjust for possible confounders, a multiple logistic regression model with the dependent variable “FLAIR-status” and independent variables “lesion volume”, “sex”, “thrombolysis” and “NIHSS” was used as a basic model (m0). In the additional model 1 (m1), “age” was added. Finally, different models were compared with regard to their ability to discriminate individuals in their FLAIR-status:i) In model 2 (m2), we added “time-from-stroke-onset” to the m1 model.ii) In model 3 (m3), we added the ROI-intensity for each rater separately to the m1 model. Paired sample statistical techniques were used for the comparison of two models. The method exploits the mathematical equivalence of the AUC to the Mann-Whitney U-statistic . The ROC curves were calculated using SPSS Statistics 21, Release Version 21.0.0.0 (SPSS, Inc., 2012, Chicago, IL, www.spss.com). The comparisons of ROC curves and the linear mixed models were done using SAS software, Version 9.3 of the SAS System for Windows. (2010 SAS Institute Inc., Cary, NC, USA).For analyzing the association between DWI-rSI and time-from-stroke-onset we calculated the mean DWI intensity over all ROIs and raters for every patient and used unadjusted and adjusted correlation analysis (Spearman's rank correlation) and a multiple linear regression analysis adjusted for “age”, “lesion volume” and “thrombolysis”. We calculated multiple linear regressions with (log-transformed) “mean DWI-value” as dependent variable and “age”, “thrombolysis” and (log-transformed) “lesion volume” as independent variables. Mean DWI intensities and lesion volume values were log-transformed to overcome the skewness in the distribution of the values. We analyzed the adjusted association between mean DWI-rSI and time-from-stroke-onset by analyzing the association of the residuals from the regression analysis with time-from-stroke-onset.Above steps were performed equally for ADC-maps. […]

Pipeline specifications

Software tools MRIcron, SPSS
Applications Miscellaneous, Magnetic resonance imaging
Organisms Homo sapiens
Diseases Cerebral Infarction, Stroke