fMRI image registration with AFNI's 3dQwarp

In the study of Parkinson's disease (PD), substantial research has shown a variety of findings within brain MR images that differ on a group-wise level between individuals with PD and healthy adults, including resting brain activity and blood flow [1, 2], diffusion tensor fractional anisotropy [3], and brain iron deposition [4, 5]. Despite these promising findings, none of these putative biomarkers have become utilized in the field. We believe this is the case principally because it has been very difficulty to register findings visible at a group level to individual brains because of the vast amount of inter-subject variability within the human population. While specific brain structures are common, the precise location, distribution, and size of these structures can vary dramatically between individuals. MR images essentially represent vast unstructured datasets. In the United States in 2011, over 8.7 million MRI brain scans sessions (each comprised of multiple brain imaging sequences) occurred. It is clear that to make practical use of this massive amount of imaging data, not only for PD but also for analysis of other conditions of the brain, high performance computing solutions will be required. There are several software packages available for image registration. For non-linear image registration, the 3dQwarp code in the Analysis of Functional Neuroimaging (AFNI) package [6] designed by NIH provides the best result in terms of normalized cross correlation, mutual information and sum of squared differences [7].

The AFNI 3dQwarp code follows the h-method in the language of the finite element analysis, which divides the image into shrinking patches and warps the source image to the template image in incremental steps. In each step, the patches are optimized in at most four sweeps (one by default). We find that for some irregular images, the results are not satisfactory, related to insufficient convergence of the solution space at each patch size. One solution to insufficient convergence uses is a linear registration step. However, our dataset (Michael J. Fox Parkinson's Progressive Markers Initiative (PPMI), a well characterized multidimensional dataset with substantial imaging data.) indicates that the linear registration in AFNI is not always well performed for PD subjects and introduces unnecessary noises to the downstream analysis. We propose a new warping procedure, which starts with an image in original space, and applies non-linear registration directly. The important part is to make sure the solution has converged at each warping step before moving to the next finer optimization step. With the new procedure, our results show about 20% improvement compared to the original procedure in term of the Pearson correlation. The practical ramifications of this improvement were apparent in the end result; before adopting this process 58 of 161 subjects were judged insufficiently converged for analysis by the research user team (Skidmore, Anthony, Marstrander above). Subsequent to the institution of the new process, while computational load increased, only 3 of 158 subjects were rejected as insufficiently registered.

Preliminary scaling studies demonstrate that the current implementation of 3dQwarp with OpenMP directives are not sufficient to take advantage of the number of cores available either on current day CPUs, or the number of cores available on the many-core Intel Xeon Phi. For both clinical and research utilization, registration code can be improved using high performance multi-thread computation. Therefore, a future direction will be to implement a version of 3dQwarp that does scale to a larger number of threads to assist in developing analytic turnaround times that will be more relevant for both research and clinical settings.