Structure-Driven Image-Warping for Anatomical Labeling of Human Brain MRI

presents a new method for the atlas-based anatomic labeling within a structure-driven deformation paradigm. The method has been tested on cranial MR images of normal healthy volunteers and multiple sclerosis (MS) patients with varying levels of atrophy. Methods A method has been developed for 3D warping of a digital brain atlas (4) to label structures in individual brain MRIs. The deformation of the atlas image is driven by anatomic correspondences between pre- segmented brain structures. For this application, the pre-segmented structures include the brain surface, mid-sagittal plane, and lateral ventricles. A complete 3D correspondence mapping is achieved through a constrained nonlinear optimization based on shape criteria that matches structures without overlaps or discontinuities. The correspondence map yields a set of displacement vectors that are interpolated to generate a full 3D deformation field for the warp. The final result is a deformed brain atlas, where the pre-segmented structures are brought into alignment directly, and the rest of the brain is deformed under constraints of continuity, effectively extrapolating local structural differences to the remainder of the volume. We tested the ability of this method to label brains with significant atrophy, as well as the applicability of the technique to longitudinal studies. A scan-rescan study was performed with three sets of MR images of 6 MS patients and 6 healthy volunteers. Images were acquired axially on a 1.5 T Siemens VISION (Siemens, Erlangen, Germany) MR imager, with a fluid-attenuated-inversion-recovery (FLAIR, TR=6000 ms, TI=2000 ms, TE=105 ms) sequence. The matrix size was 256x256 with a 230 mm field of view and 5mm slice thickness. Thirty interleaved slices were acquired (full gap and fill) for complete coverage of the brain. Each subject left the scanner between consecutive scans, emulating a "real-life" error, as it may apply for serial studies. The images were labeled using the new atlas warping method, and volumes were calculated for labeled structures in each image. The error for labeling individual structures was calculated as the coefficient of variation (COV) between the 3 repeated measurements.