To reduce operator-induced variability in manual selection of the VOI for MRS, several automated methods were proposed^1~3. However, they were mainly focused on the VOI re-positioning for the follow up scans using rigid registrations. We wanted to create a tool that would work in subjects with varying head sizes and anatomies, thus accommodating the head size normalization and anatomical irregularities. We hypothesized that non-rigid registrations will minimize inter-subject variability in automatic VOI placements. Hence, we tested if affine and nonlinear b-spline registrations would improve between-subject reproducibility of VOI positioning relative to manual and automated positioning using rigid registration.

The implemented scheme utilizes the 3D anatomical data of the subject and runs the registration of the atlas brain to the subject brain to estimate the VOI position over the subject brain (Fig 1). The target VOI were defined over T1 1mm MNI-152 brain atlas⁴. In addition, VOIs were placed manually on subjects to compare to the automated placements. Two VOIs were defined on the Posterior Cingulate Cortex (PCC, 20x20x20mm³) and Left Hippocampus (LHC, 13x12x26mm³) using anatomical landmarks employed by previous studies^5,6. Datasets obtained from 12 subjects (4 females, mean age 31±8) enrolled at two sites (Center for Magnetic Resonance Research, CMRR, University of Minnesota; Korea Advanced Institute of Science and Technology, KAIST) were used. 7 datasets from CMRR were acquired using a 3T Siemens Trio, and 5 datasets from KAIST were taken using a 3T Siemens Verio. Both sites acquired T1-weighted 3D MPRAGE images at 1x1x1mm³ resolution with 32Ch head coil (TR/TI/TE=2530/1100/3.65ms, 176x224x256mm³ FOV for CMRR; TR/TI/TE=1900/900/2.48ms, 176x256x256mm³ FOV for KAIST).

FSL-BET⁷ was used with robust setting for brain extraction. For the registration and transformation, we used Elastix and Transformix tools⁸, based on the Insight Toolkit (ITK) library⁹. Rigid, affine and b-spline registrations were optimized at a fixed iteration number defined beforehand: 200 iterations each for rotation and translation, and 500 iterations each for affine and b-spline registrations.

All calculations were performed on a desktop PC (Fedora 21, i5-4670@3.40Ghz, 8GB-RAM, 1TB-HDD, NVIDIA GeForce GTX650 1GB-VRAM). The variability in the head volume across subjects was measured to understand the need for VOI volume normalizations. T1 intensity values within the VOI masks were also extracted to estimate White Matter (WM), Grey Matter (GM) and Cerebrospinal Fluid (CSF) content of each voxel using Gaussian Mixture Model (GMM) methods. Finally, the transformed VOI masks were registered back to the MNI-152 atlas to evaluate the overall spatial variability of the positioning across the subjects.

The average calculation times for rigid, affine and b-spline registrations were 24.50(±1.47), 33.95(±1.60), and 63.87(±2.81) seconds respectively. Mean head volume was 1.56 liters with coefficient of variance (CV) of 10.4%, suggesting that variations in the head volumes across subjects would affect the VOI selection consistency. Note that while rigid registrations cannot handle volume normalizations, affine and b-spline registrations can.

For WM/GM/CSF estimations, the lowest CV was obtained with b-spline registrations (Fig 2), confirming lower variance in VOI tissue content than manual positioning and automated VOI selection using rigid and affine registrations.

To analyze spatial variations in VOI selection, each binary VOI mask of 12 datasets per positioning method were added; then the fraction of the VOI that is included in all 12 VOI (# of pixels in all 12 sets divided by total # of pixels in the VOI) was computed. This analysis showed that the spatial variation was the least when using the b-spline registration, with the greatest percentage of overlapped volumes out of all registrations in the PCC and LHC (Fig 3 and 4).

The results show that b-spline registration yields the most consistent registration across subjects, by yielding the greatest spatial overlap percentages and lowest CV values for the composition (WM/GM/CSF) intensity variations in VOI placements. While rigid registration may offer improvements in VOI placement reproducibility over manual placement (Fig. 3), it is not sufficient for the anatomic variability that exists between subjects. Among non-rigid registrations, the benefits of greater performance with b-spline over affine appear worth the increased time of 30 seconds, which could further be reduced by leveraging improved algorithms and better hardware such as GPU. Hence, for registration-based automatic VOI positioning in multi-subject MRS studies, nonlinear registrations such as b-spline must be considered for reliable outcomes.

This study demonstrates the feasibility of a reliable cross-subject automatic VOI positioning, which could lead to increased clinical adoption of MRS by reducing operator dependence and making it easier for MR technologists to acquire high quality MRS data.