On Statistical Analysis of Neuroimages with Imperfect Registration

摘要: A variety of studies in neuroscience/neuroimaging seek to perform statistical inference on the acquired brain image scans for diagnosis as well understanding pathological manifestation diseases. To do so, an important first step is register (or co-register) all data into a common coordinate system. This permits meaningful comparison intensities at each voxel across groups (e.g., diseased versus healthy) evaluate effects disease and/or use machine learning algorithms subsequent step. But errors underlying registration make this problematic, they either decrease power or follow-up tasks less effective/accurate. In paper, we derive novel algorithm which offers immunity local deformation field obtained from procedures. By deriving invariant representation image, downstream analysis can be made more robust if one had access (hypothetical) far superior procedure. Our based recent work Scattering coefficients. Using starting point, show how results harmonic (especially, non-Euclidean wavelets) yields strategies designing and additive noise representations large 3-D volumes. We present set synthetic real images where achieve even presence substantial errors, here, standard procedures significantly under-perform fail identify true signal.