111 Etiology in Lesion-Symptom Mapping: Tumor vs. Stroke or radiologist was consulted. The interrater reliability was calculated based on eight stroke lesion masks using the same method as for the tumor interrater reliability. The interrater reliability for all three raters was 81.3% (range 69.8-91.1%), as reported previously.29 Pre-processing Each individual’s FLAIR and binary lesion mask (both tumor and stroke) were normalized to an age-specific older adult MNI template using the Clinical Toolbox35,36 implemented in SPM12 (www.fil.ion.ucl.ac.uk/spm). For unilateral lesions, enantiomorphic normalization was applied to reduce distortions in the normalization due to the lesion.37 For bilateral lesions normalization with cost function, masking was applied.38 If the normalization results using enantiomorphic normalization was unsatisfactory for the tumor data, the normalization process was repeated using cost-function masking. The superior lesion mask (defined as visually best representing the original lesion location) was used in subsequent analyses. After spatial normalization, each lesion mask in MNI space was visually compared to the lesion in native space, and manually corrected if needed. Subsequently, the lesion mask was smoothed with a Gaussian kernel of 3 mm at FWHM. For stroke patients, normalization was optimized for patients with enlarged ventricles (>1.5 SD above ventricle volume in elderly template) using a warping regularization reduced by one order of magnitude. The resolution of the normalized lesion maps was 2x2x2 mm3. All results are displayed in neurological orientation (left = left hemisphere). Data analysis Multivariate lesion-symptom mapping LSM was applied to test the relationship between lesion status in each voxel and cognitive performance, defined as a Z-score, for each task. For the multivariate LSM we used the support vector regression LSM (SVR-LSM) toolbox running under Matlab2019a (The MathWorks, Inc., Natick, Massachusetts, United States), which is a multivariate regression algorithm based on machine learning (github. com/atdemarco/svrlsmgui).39,40 This multivariate method, as opposed to a massunivariate approach, considers intervoxel correlations and therefore is potentially more sensitive to examine lesion-behavior relationships.39,41 It has been successfully used and validated in multiple studies including both real and simulated lesion data.e.g.1,22,39,42 As no clear criteria on parameter choice are available yet, the hyperparameter values were kept at a cost of 30 and a gamma of 5, following the original paper.41 A nonlinear radial basis function kernel was used. A lesion threshold of 3 (i.e. at least 3 patients with lesioned tissue at each voxel) was applied. To test the significance of the beta values, permutation testing was used 5
RkJQdWJsaXNoZXIy MTk4NDMw