Spatial Registration of Neuroimaging Data: Analysis of the Convenience of Performing Non-Affine Transformations

Computer-based analysis of neuroimaging data in multisubject studies requires a previous spatial registration procedure, which ensures that the same voxel across different images refers to the same anatomical position. Several algorithms have been proposed to this end and most of them perform the spatial registration in two steps, an affine transformation followed by a non-linear registration. While the former applies only translations, rotations, zoom and shears to the neuroimages, the non-linear registration step can deform them to adjust the size and shape of individual regions. Although the scientific community generally accepts that these transformations are necessary, even though they may introduce certain distortions (noise), some recent works indicate that it is preferable to perform the spatial registration as an affine transformation only, in order to prevent the non-linear registration from removing information that could be relevant in the further analysis. In this work we evaluated the influence of applying nonlinear transformations during the special registration of molecular neuroimages that will be used in computer systems intended to assist the diagnosis of neurodegenerative disorders. Specifically, we compared the performance of a Support Vector Machine classifier that used data spatially registered using only affine transformations and other one that used data that have been registered using the classical procedure, which includes non-linear transformations. Two datasets were considered, one intended to assist the diagnosis of Alzheimer's disease and other one intended to assist the diagnosis of Parkinsonism. The results suggest that non-linear transformations facilitate the subsequent classification and provide slightly higher accuracy rates. The different is more important with data in which the intensity is concentrated in a small target region such as DatSCAN neuroimages, used to assist the diagnosis of Parkinsonism.