IC‐P‐145: Image processing pipeline for deformable registration
Leonid Teverovsky, Oscar L. López, Howard Jay Aizenstein, Meryl A. Butters, Julie C. Price, James T. Becker · Alzheimer s & Dementia · 2011
We present a fully automated, distributed pipeline that utilizes publicly available and in-house software. The pipeline performs image conversion, skull stripping, bias correction, cross-subject intensity normalization, midsagittal plane alignment, rigid, affine and fully deformable registrations, computation of dataset-specific template, grey/white/CSF tissue segmentation, atlas-based segmentation (ABS) and volumetric measurements of user-defined regions of interest (ROI). The pipeline can perform reference-to-input and input-to-reference image registration and handle longitudinal datasets. Skull stripping uses Brain Extraction Tool (BET), mri_watershed, or deformable registration, as selected by the user. Bias correction uses inormalize and N3 tools of MINC suite. Cross-subject intensity normalization is implemented using ITK. Midsagittal plane alignment uses in-house software. Rigid and affine registrations use flirt tool of FSL. Deformable registration uses fnirt of FSL or finite-element mesh registration followed by Demons registration implemented using ITK. Grey/white/CSF tissue segmentation utilizes fastor ABS. Both ABS and deformable registration-based skull stripping are done by registering the reference brain with delineated ROIs to the input brain and transferring ROI labels. We tested the pipeline using data acquired by the University of Pittsburgh's Alzheimer Disease Research Center (ADRC) over 15 years, on 1.5T and 3T scanners of different manufacturers and models using 3D MPRAGE and 3D SPGR protocols. ADRC dataset consists of 717 images of both genders over age 50. We processed the scans using the following steps: image conversion, skull stripping via deformable registration, bias correction, intensity normalization, affine and deformable registration. The colin27 template was the reference image. We did not adjust parameters for any subset of images, and the pipeline did not fail on any of the 717. Processing took approximately 100 hours on three Intel Quad Core2.6Ghz, 8Gb RAM desktops connected by 10Mbs LAN. We proposed a fully automated distributed image processing pipeline. Our preliminary experiments with the pipeline using a large heterogeneous dataset demonstrate that the pipeline is scalable, robust and does not require parameter tweaking for different image subsets. Representative pipeline output