P6. Toward optimized nonlinear deformation algorithms for subcortical DBS target regions in human brain MRI

Abstract

Objective 1. To evaluate and optimize commonly used nonlinear deformation algorithms when applied to human brain MRI. 2. To evaluate and optimize accuracy of automated atlas-based segmentation of the two most commonly used target regions for deep brain stimulation (DBS), the nucleus subthalamicus (STN) and the internal part of the globus pallidus (GPi). 3. To guide clinicians as to which preoperative MRI sequences to acquire in patients undergoing DBS. Background Translating single-subject imaging into a common reference frame such as the MNI space is a ‘core concept within the field of brain mapping’ (Evans et al., 2012) making an accurate transformation into this space necessary. Here, we evaluated and optimized 6 commonly used deformation algorithms and compared each outcome to manually labeled brain regions of 103 brains. Methods Algorithms evaluated were: New Segment, DARTEL and SHOOT (SPM); FNIRT (FSL); SyN and BSpline (ANTs). Target space was the MNI 2009b NLIN space. An atlas of the STN and GPi in MNI space (Ewert et al., 2017) was transformed to native space by inverting each deformation matrix. The resulting atlas-based segmentations were then compared to expert manual segmentations of the native brains. Overlap between the two was quantified using the Dice coefficient (Dice et al., 1945), mean surface distance and correlation of volumes. Two open source datasets were used: The IXI-dataset ( http://www.brain-development.org/ixi-dataset/ ) and the HCP-dataset ( http://www.humanconnectomeproject.org/ ). Results The best performing algorithms were New Segment with a custom tissue probability map (TPM) (1) and ANTs SyN (2) with an additional step to refine subcortical coregistration. Performance is compared to inter-rater results (3) of manual segmentations. Fig. 1 shows results for the HCP data for STN (L, in red) and GPi (R, in blue) separately: median Dice for STN (1): 0.70 (std: 0.1); (2): 0.72 (0.06); (3): 0.76 (0.09). Median of mean surface distance in mm for STN: (1): 0.45 (0.14); (2): 0.38 (0.07); (3): 0.41 (0.2). Correlation between automated and manual volumes are for (1) R = 0.296, p . Conclusion First, we identified two nonlinear deformation algorithms that perform superiorly to other commonly used algorithms in the field. We then further optimized their performance resulting in a more precise alignment of subcortical structures. The results of atlas-based segmentations are similar to inter-rater results made by expert manual raters. Second, multimodal image registration was more accurate than unimodal image registration. We conclude that preoperative imaging of DBS patients should include at least a T1- and T2-weighted modality.