Mapping internal brainstem structures using T1 and T2 weighted 3T images

Abstract

Background: Many neurodegenerative diseases affect the brainstem and often do so in an early stage. The overall goal of this project was (a) to develop a method to segment internal brainstem structures from T1 and T2 weighted sequences by taking advantage of the superior myelin contrast of the T1/T2 ratio image (RATIO) and (b) to test if this approach provides biological meaningful information by investigating the effects of aging on different brainstem gray matter structures.

Methods: 675 T1 and T2 weighted images were obtained from the Human Connectome Project Aging. The intensities of the T1 and T2 images were re-scaled and RATIO images calculated. The brainstem was isolated and k-means clustering used to identify five intensity clusters. Non-linear diffeomorphic mapping was used to warp the five intensity clusters in subject space into a common space to generate probabilistic group averages/priors that were used to inform the final probabilistic segmentations at the single subject level. The five clusters corresponded to five brainstem tissue types (two gray matters, two mixed gray/white, and 1 csf/tissue partial volume).

Results: These cluster maps were used to calculate Jacobian determinant maps and the mean Jacobians of 48 brainstem gray matter structures extracted. Significant linear or quadratic age effects were found for all but five structures.

Conclusions: These findings suggest that it is possible to obtain a biologically meaningful segmentation of internal brainstem structures from T1 and T2 weighted sequences using a fully automated segmentation procedure.

Keywords: T1; T2; aging; brainstem; internal structures; segmentation.

Figure 1

Summary of brainstem segmentation pipeline. The pipeline consists of 3 main modules of which each encompasses several steps. The first module is “pre-processing” that uses SPM12 routines for tissue segmentation with inbuilt additional bias correction and brainmask generation, followed by spatial normalization to the MNI space while maintaining the original image resolution. The T1 and T2 image are re-scaled and the T1/T2 or RATIO image calculated. The rois used to extract the gray (blue) and white (red) matter intensities reported in Table 2 are shown in the insert. The images are then passed on to the second module whose first step is to use a binary brainstem/thalamus mask in MNI space to extract the brainstem/diencephalon from each of the three images. The next step uses a k-mean clustering algorithm to identify 5 intensity clusters. The cluster labels are converted into an image in subject space as binary first pass segmentations. This is followed by the generation of a group average probability map or prior map for each cluster by warping the first pass binary segmentations into a common space using SPM's DARTEL “create template algorithm” which is also the first step of the last module or “final segmentation,” i.e., the generation of probabilistic group averages to be used as priors to refine the segmentation outputs. The transformation matrix from this step was inverted and used to warp the probabilistic group averages into the subject/MNI space. The information from the priors was combined with the distance information from the clustering step which allowed to clean-up voxels assigned to a cluster not consistent with the probability information and to convert the binary first pass segmentation into a probabilistic final segmentation. Please see “Methods” for more details.

Figure 2

Brainstem segmentation outputs. The first 5 left-sided rows are 3D renderings and the corresponding projections of the 5 probabilistic population cluster maps or priors onto a population T1 brainstem image that were generated with DARTEL (A) The first 5 right sided rows show the final segmentations of an individual subject (B) The row below depicts a 3D rendering of the brainstem labels and the labels overlaid on the population T2 brainstem image (C) Below (D) is a Jacobian Determinant map generated by the standard VBM approach (input whole brain gray and white tissue maps) whereas the bottom row (E) shows the Jacobian Determinant map generated by supplying the 5 final segmentations of an individual to DARTEL. Details of individual structures, e.g. RN, NTS etc. are identifiable.

Figure 3

Brainstem structures with significant age effect after correction for multiple comparisons with (FDR, q = 0.05). The roi color reflects the regression coefficient strength (please see also Table 3), negative coefficients are indicated in cold and positive coefficients in warm colors. CI, colliculus inferior; CR, ncl. reticularis cuneiformis; CS, colliculus superior; LC, locus coeruleus; NR, ncl. ruber; NTS, ncl. tractus solitarii; OI, ncl. olivarius interior; OR, ncl. raphe obscurus; PAG, periaqueductal gray; PN, pontine nuclei; PR, ncl. raphe pallidus; RPC, ncl., reticularis pontis caudalis; RPO, ncl. reticularis pontis oralis, RPT, ncl. reticularis pontis tegmenti; SC, ncl. subcoeruleus; SN, substantia nigra; Tld, ncl. tegmentalis laterodorsalis; Trm, ncl. tegmentalis rostromedialis; VLM, ventrolateral medulla; VTA, ventral tegmental area.