36 Chapter 2 transformations. The nonlinear DCT consisted of 16 discrete-cosine basis functions along each dimension.20 For all the above-mentioned transformations, all registration steps were combined in a single joint transformation with a single interpolation from native to standard space. Quantitative Evaluation We compared the four normalization methods quantitatively by studying the overlap between the individual CBF image registered to the standard space and a CBF template in the standard space using the Tanimoto similarity coefficient for real-valued vectors, see equation (1) in the work of Anastasiu et al.21 The Tanimoto coefficient (TC) is a measure of image overlap, ranging from 0% (completely dissimilar) to 100% (identical images). If we assume that perfect registration does not lead to identical images but ones that still retain physiological differences, TC >70% can be regarded as excellent image agreement. We masked the brain in both images and normalized the values in each image to the 97th percentile value while excluding the higher signal values from the computation. Note that we computed the whole-brain TC for continuous perfusionweighted values (also known as Tanimoto distance) rather than the commonly used measure for binary images. 21 Qualitative Evaluation The spatial normalization of ASL images to the MNI space was inspected qualitatively by examining the overlay of CBF and outer WM borders in MNI space with a threshold at 50% WM partial volume. The T1w-based and the three types of ASL-based normalizations were visually scored by three raters: a pediatric neuroradiologist (MD, 7 years of radiology experience), an ASL image processing engineer (JP, 10 years of experience), and a neuroradiologist (VK, 10 years of experience). The visual alignment of the ASL images in MNI space was categorized into four quality categories: (1) unusable, 2) poor, (3) usable, and (4) excellent. For the evaluation of the results of each normalization method of each subject, the rater rated overviews of normalizations showing 12 axial and 12 sagittal slices. All visual overlay images from both the patient and control groups for all four normalization methods were pseudorandomized and pseudonymized. The pseudonymwas visible to the raters throughout the whole procedure of individual rating and for reaching consensus. The raters were blinded to the method and clinical history of the patients and controls.
RkJQdWJsaXNoZXIy MTk4NDMw