Spring 2016
Research Project 1, 26 ECTS
Atlasing the Human
Subcortex:
Sequence development
Nicholas Judd 11118032
treated in extreme cases via subcortical deep brain stimulation (DBS). Yet current magnetic resonance (MR) atlases lack the anatomical precision needed to localize these subcortical nuclei of interest, highlighting the need of a
comprehensive subcortical MR atlas. Using high-field (7T) MR technology, it is now possible to visualize more of these previously unattainable nuclei. The goal of this project is two-folded;; firstly to develop an optimal MR sequence. The second goal is to segment the striatum. These two goals are needed for the grander ambition, which is to obtain visualization of as many subcortical nuclei, in multiple modalities, sufficient for segmentation. This is a prerequisite towards building a comprehensive subcortical MR atlas. Firstly we report our signal-to- noise and contrast-to-noise calculations which led us to develop a novel MR sequence;; Magnetization-Prepared 2 Rapid Gradient Echo with Multiple Echoes (MP2RAGEME). MP2RAGEME is a multiple contrast MR-sequence that allows the visualization of multiple subcortical nuclei in different MR-modalities. For the segmentation of the striatum a guideline was developed and three main
problems were identified;; The Accumbens Nucleus (NAcc) problem, Striatal cell bridges problem and the islands or noise problem. Inter-rater reliability (IRR) of segmentations, in the form of Dice coefficient and the modified Hausdorff
distance, are reported and generally very high. These measurements will also be used as a secondary measure to SNR/CNR for determining sequence efficacy. Primarily this result, of high IRR measurements, lead us to value the merit of a segmentation guideline for the striatum.
Introduction
Parkinson’s disease, dystonia and essential tremors are a result of neural deficits within the thalamus and basal ganglia (Hariz, Blomstedt, & Zrinzo, 2013).
Recently, it has been shown that the application of Deep Brain Stimulation (DBS) helps to reduce symptoms in a range of neurological disorders, most commonly Parkinson’s. DBS is a surgical procedure, in which surgeons plant electrodes to target nuclei of interest, dependent upon the disorder, for electrical stimulation (Trépanier, Kumar, Lozano, Lang, & Saint-Cyr, 2000). Partially due to the successful deployment of DBS to treat Parkinson’s, the field has advanced, to study a myriad of potential neurological and psychiatric disorders;; including, but not limited to, pain, epilepsy, obsessive compulsive disorder, depression, bipolar disorder, autism, post traumatic stress disorder, eating disorders, addiction, cognition and tinnitus (Hariz et al., 2013). Each disorder has a corresponding anatomical pathway implicated, therefore leading to disorder specific structures as potential targets for DBS. For example, in patients with treatment resistant (to electroconvulsive therapy & antidepressant medication) depression, it has been shown that stimulation of the ventral striatum reduces symptoms (Malone et al., 2009).
Noninvasive visualization of these DBS targets is crucial for a precise localization of the DBS electrodes;; yet magnetic resonance imaging (MRI) atlases lag behind in defining the boundaries of these targets. Alkemade and colleagues in 2013, highlight the lack of anatomical structures in MRI atlases. This was accomplished by comparing the defined subcortical structures in the Federative Community of Anatomical Terminology to those listed in current MRI atlases, resulting in the finding that only 7% of named subcortical gray matter structures are accounted for. One possible reason for this large discrepancy, between named structures and those available in MRI atlases, is due to the low spatial resolution of current MRI technology. Recent advances in the imaging field, such as high-field (7T) MRI, possess the ability to reveal previously unattainable subcortical structures. The advantages of a 7T scanner are the increase of signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), yet one main disadvantage is the high cost. Therefore its use is only warranted in cases when the increase of resolution is an absolute necessity, for example to visualize small subcortical structures (Webb & Van de Moortele, 2015). Aptly, multiple recent review papers on DBS highlight the utility of high-field MRI to distinguish and accurately target nuclei of interest (Hariz et al., 2013;; Kopell & Greenberg, 2008;; Udupa & Chen, 2015). Adding evidence for the necessity, particularly within the DBS field, of a subcortical high- field MRI atlas.
interest. Here, we will develop an MR sequence with sufficient visualization to segment as many subcortical structures possible;; these segmentations will lead to the development of a comprehensive in-vivo MR atlas. Firstly, I will briefly review the MR sequences shown to produce high enough visualization to be used for manual segmentation of subcortical structures, such as Magnetization- Prepared Rapid Gradient Echo (MPRAGE), Magnetization-Prepared 2 Rapid Gradient Echo (MP2RAGE) and Quantitative Susceptibility Mapping (QSM) (de Hollander et al., 2014;; Tourdias, Saranathan, Levesque, Su, & Rutt, 2014). This report will then proceed to explain, and later on, show the results of the methods used for determining the optimal scan sequence, such as SNR/CNR calculations and inter-rater segmentation comparisons. A second aspect of this report is the segmentation of the striatum, the methods and results of which will be presented and later on discussed. Finally, the report will conclude by discussing the factors and considerations leading to the optimized scan sequence and its utility towards building a subcortical MR atlas.
MR sequences of interest
The main MR sequences of interest for this report are MPRAGE, MP2RAGE and QSM. Below I will briefly explain the characteristics of these sequences along with their utility for subcortical visualization.
MPRAGE
Tourdias and colleagues (2014) demonstrate how MRI sequences, such as MPRAGE, can be optimized to visualize intra-thalamic nuclei. The authors used a high-resolution MPRAGE sequence with a resolution of 1mm isotropic. The MPRAGE sequence provides a T1-weighted scan with minimal specific
absorption rates (SAR) and relatively short scan times (6.8 min). Typically the inversion rate is set to null the cerebral spinal fluid (CSF), resulting in increased gray to white matter contrast (Saranathan, Tourdias, Bayram, Ghanouni, & Rutt, 2015). Tourdias and colleagues article aptly demonstrates how scan parameters, such as inversion time (670ms), can have profound effects on visualization even if the timing only differs by +/- 50ms (Figure 1). Delving into the signal equation for MPRAGE reveals multiple intertwined parameters;; such as inversion pulse repetition time (TS), number of readouts per inversion (N), excitation flip angle
(α), readout sequence repetition time (TR), receiver bandwidth (BW) and RF pulse length. Saranathan and colleagues (2015) state these scan parameters have a direct effect on SNR and blurring characteristics of the image.
Optimization of these intertwined parameters is critical to increase SNR and CNR, which in turn allows a better segmentation of any structure of interest. The researchers conclude by determining a white matter nulled (WMn) scan, with a flip angle of 4 degrees and a TS of 6s, produces the best image for intra-thalamic delineation in the shortest amount of time (5 mins).
Figure 1: Contrast changes with differing first inversions (Tourdias et al., 2014, p. 540)
MP2RAGE
MP2RAGE is another T1-weighted scan, which is a modified version of MPRAGE
using two images at different inversion time points (Figure 2/3/4). This sequence has shown promise in 7T to visualize subcortical structures, such as the striatum (Keuken et al., 2014;; Marques et al., 2010). Combining these two images allows for a single image free of T2* contrast, proton density contrast and reception bias
field (Marques et al., 2010). MP2RAGE tackles one of the greatest issues hindering 7T scanning, transmit field inhomogeneities (Jose P. Marques & Gruetter, 2013). The end result is an anatomical image with higher SNR/CNR allowing for visualization and segmentation of these elusive subcortical
weighted imaging as a superior alternative. Quantitative susceptibility mapping (QSM) is a T2*-weighted MR calculation that has shown promise for the
visualization and segmentation of subcortical nuclei (Darki, Nemmi, Möller,
Sitnikov, & Klingberg, 2016;; Keuken et al., 2014). QSM utilizes raw gradient-echo phase data to visualize magnetic susceptibility (Schweser, Deistung, Sommer, & Reichenbach, 2013). Therefore it is particularly useful to image iron rich tissue, such as the medial medullary lamina (Figure 2), which acts as a border to separate the internal from the external segment of the Globus Pallidus (Gp) (Keuken et al., 2014). The same research group used QSM to examine the borders of the STN;; hypothesizing if distinct anatomical borders are present the iron content will reflect this characteristic (de Hollander et al., 2014). The iron content indicated a gradual increase towards the medial-inferior tip, adding evidence against the view of distinct anatomical borders.
Figure 2: Medial Medullary Lamina (Max et al., 2014, p. 41)
Multimodal Approach
Combining these differing scan modalities to segment subcortical structures offers the most promising route to build a MRI atlas of the subcortex. Keuken and researchers (2014) aptly demonstrated this by developing a multimodal atlas, derived from 30 young subjects, of six subcortical structures: the striatum, the GPi/e, the substania nigra (SN), the subthalamic nucleus (STN) and the red nucleus. The sequences used where a whole brain MP2RAGE (t = 10:57 mins;; voxel size=0.7mm isotropic), a zoomed MP2RAGE (t = 9:08 mins;; voxel
size=0.6mm isotropic) and a muli-echo 3D FLASH (t = 17:18;; voxel size=0.5 mm isotropic). Using the acquired FLASH data QSM maps were derived, adding another contrast modality (Figure 2). In collaboration with the FSL research group, an automated multimodal segmentation tool was developed, aptly named MIST (Mulitmodal Image Segmentation Tool). MIST has the ability, when utilizing multiple scans, to produce high quality segmentations of the striatum (Visser et al., 2016).
Optimizing the MR sequences
The primary aim of this research project is to develop an MP(2)RAGE based MR sequence, which results in the visualization of as many subcortical structures feasible for segmentation. When developing an MR sequence it is important to consider the tradeoff between time, resolution and signal.
Since this sequence will eventually be utilized in Parkinson’s patients acquisition time should be minimized. To this end, one might want to use acceleration
techniques that minimize acquisition time. For example, sense is a MR technique that utilizes multiple receiver coils in parallel to drastically reduce scan time (Pruessmann, Weiger, Scheidegger, & Boesiger, 1999). Since acquisition time is limited, sense may be a necessity for our sequence. However, if utilized it will be at the expense of MR signal.
To examine this tradeoff five key questions, listed below, will be calculated for SNR and CNR. SNR and CNR are useful quantities for image evaluation and contrast enhancement (Dietrich et al., 2007).
1) What is the ideal resolution (0.7 isotropic vs. 0.5 isotropic) for visualization of subcortical structures?
5) What is the contrast difference between MP2RAGE &
MP2RAGE with four echoes (MP4/MP2RAGEME)?
Once the optimal sequence has been determined;; segmentations from independent raters will be compared using quantitative measurements. These scores will act as a secondary measure to SNR/CNR for determining the efficacy of the MR sequence
developed. Yet prior to segmentation a detailed guideline must be developed;; this is necessary to facilitate consistency across raters. Due to time limitations this project will focus on segmentation of the striatum, specifically the Putamen (Pu), Caudate (Cd) and Nucleus
Accumbens (Ac). The quality of both the scan and the segmentation guideline can be indirectly measured via inter-rater reliability of segmentations. Two independent raters will use the same scans and segmentation guidelines. The final step will be to compare our
segmentations with the automated tool MIST.
Segmentation of the Striatum
Manual segmentation is the process in which two independent raters outline the boarders of a structure of interest in a continuous direction in space. Manual segmentation is still advantageous to automatic segmentation primarily since the tools have not yet been comprehensively developed. As briefly mentioned MIST
is an automatic subcortical segmentation tool, yet it only includes six subcortical structures and uses a reference mesh from a manual segmentation (Visser et al., 2016). One of the main limitations of manual segmentation is the time in which it takes to complete. Yet this processes is a necessity in the development of a subcortical atlas. As aforementioned, the main justification for building a subcortical atlas is to assist surgeons in positioning DBS electrodes.
The above suggests the need for a comprehensive subcortical MR atlas, along with outlining promising MR sequences for visualization. The goal of this study is the development of an MR sequence that visualizes as many subcortical
structures possible for segmentation;; these segmented structures will eventually be used to develop an atlas. The following section of the report will outline the methods used to reach this twofold goal;; MR sequence development &
segmentation of the striatum.
Methods
As seen in the project overview this project is divided into two sections;; 1) optimizing MR sequences and 2) segmentation of the striatum. Firstly the
methods used to determine the optimal MR sequence (SNR & CNR calculations) will be described. The second subsection, segmentation of the striatum, will outline the methods used to for inter-rater reliability comparisons.
Optimizing MR sequences
To acquire the structural data four participants were scanned in a variety of sequences.
Participants
All subjects signed informed consent documents prior to scanning. Monetary compensation (20e) was granted to three subjects, one was excluded since they were university staff. All structural scans were acquired on a Philips 7T Achieva MRI with a 2-channel transmit, 32-channel receive head coil from Nova Medical (NOVA Medical Inc., Wilmington MA) at the Spinoza Center for Neuroimaging. Since this project is the advancement of a technique, multiple differing MR sequences were used, seen on table one below. The voxels in all sequences were acquired isotropic yet reconstructed slightly non-isotropic, the reconstructed voxel sizes can be found in Appendix A.
Table One MR-Sequence Overview
Scan Type Voxel Slices Time TR TE Inv1 Inv2 flip 1 MP2RAGE 0.5mm 328 13:00 6000 2.4 680 1965.5 7 7 2 MP2RAGE 0.7mm 234 13:00 6000 3 680 1965.5 4 4 3 MP_nosense 1mm 164 16:00 6000 3 670 4 4 MP_sense 1mm 164 6:30 6000 3 670 4 5 MP2RAGE 0.7mm 234 13:00 6000 3 670 1965.5 4 4 6 MP2RAGE 0.7mm 234 13:00 6000 3 670 1965.5 4 4 7 MP2RAGEME 0.7mm 234 14:35 6700 3/8.9/14.8/20.7 670 3749 4 4 8 MP4SLAB 0.6mm 128 9:04 6700 2.3/8.2/14.1/20 670 3749 4 4 9 MP2RAGEME 0.7mm 234 20:42 6868 3/11.5/19/27.5 670 3674 4 4 TR Time of Repetition, TE Time of Echo, Inv1 Inversion 1, Inv2 Inversion 2, flip Flip Angle
Signal-to-noise ratio (SNR)
We define signal-to-noise ratio as “the relative strength of a signal compared with other sources of variability in the data” (Huettel, Song & McCarthy, 2015, p. 273). Meanroi SDroi Contrast-to-noise ratio (CNR)
We define contrast-to-noise ratio as “the magnitude of the intensity difference between different quantities divided by the variability in their measurements” (Huettel, Song & McCarthy, 2015, p. 273). Since our interest lay in segmenting subcortical nuclei, our equation focused on gray to white matter contrast
differences.
Meanroi+ MeanWM
SDWM
Regions of interest (ROI) for SNR/CNR calculations
Spheres were built in individual space using command line code of fslmaths and fslstats (FSL 5.0.9). When comparing different sequences within the same
Tool (FLIRT) with six degrees of freedom. The initial goal was to have an equal number of voxels in each sphere across resolutions, yet this is not possible due to the mathematical properties of a perfect sphere. Appendix A lists the amount of voxels and volume for each sphere in differing resolutions along with their respective radii. The regions of interest were White Matter (WM) in the Corpus Callosum (cc), Putamen (Pu), Caudate (Cd), Cerebral Spinal Fluid (CSF) in the Lateral ventricle and the medial dorsal thalamic nucleus (MD). To dampen the inhomogeneity effects inherent to 7T each ROI had two spheres placed
equidistant on both hemispheres. The size of the sphere was limited by two contingent factors: the structure with a limiting dimension in space (cc) in the largest voxel size (1mm).
Segmentation of the striatum
Two independent raters (Katja Höhne & Nicholas Judd) segmented six scans (1, 2, 5-8), strictly following the guidelines above, using FSL (5.0.9). This resulted in 24 masks, which were then checked and converted into binary masks.
Inter-rater Reliability (IRR)
The goal of measuring inter-rater reliability is to determine if there is consistency across raters. As aforementioned, this project is unconventional by using the IRR to indirectly measuring the quality of both the MR sequence’s and the
segmentation guidelines. MATLAB (version 2015b, The MathWorks, Natick, MA) on Mac Yosemite (10.10.5) was used to compute the two statistical comparison methods used;; Dice coefficient and the modified Hausdorff distance (MHD) (Beauchemin, Thomson, & Edwards, 1998;; Dice, 1945). The Dice coefficient tends to be biased when comparing masks with larger surface areas, since naturally they will have more matching surface space than a smaller mask. This limitation of Dice coefficient is precisely why we decided to also report the MHD;; since its “value increases monotonically as the amount of difference between the two sets of edge points increase” (Dubuisson & Jain, 1994, p. 568). Dubuisson and Jain (1994) empirically show the MHD to be more robust to outlier points then the Hausdorff distance, leading us to use the MHD. Scripts will be provided upon request.
MIST comparison
MIST has yet to be released by FSL, therefore the MIST comparison part of this project was not possible.
Results
The results section will firstly focus on the MR sequences as a whole before moving on to the results surrounding the development of the segmentation protocol and the inter-rater reliability scores.
Magnetization-Prepared 2 Rapid Gradient Echo with Multiple Echoes (MP2RAGEME;; MP4)
One of the more promising results of this project was the invention of the MP2RAGEME sequence. The sequence is an extension of an MP2RAGE, in which there is a first inversion followed by four echoes. The first echo acts as the second inversion for field inhomogeneity correction. One of the benefits of this sequence is the ability to also derive QSM maps from the same sequence.
Scan Quality
1) What is the ideal resolution (0.7 isotropic vs. 0.5 isotropic) for visualization of subcortical structures?
2) Does a change in the first inversion to match Tourdias et al. (2014) paper affect CNR?
Table 2 Contrast to Noise Ratio
comparing differing first inversions
Scan 1 (0.5mm) Scan 2 Scan 5 (Tour) Cd -4.5 -10.7 -10.8 Pu -5.6 -9.7 -10.1 Cd Caudate, Pu Putamen, Tour Tourdias et al. (2014)
As seen in Table 2, a 0.5mm whole brain MP2RAGE contains too much noise to be useful. Figure 3 shows the comparison between scan 1 and scan 2. The scan with the first inversion to match Tourdias and colleagues (2014) paper (scan 5)
shows slightly better CNR than the scan with a 10ms longer first inversion. This result should be interpreted with caution due to the limited amount of sequences acquired (1 sequence directly comparable) and the minimal contrast difference calculated (Cd = -0.01, Pu = -0.04). SNR calculations of these scans are in Appendix B.
Figure 3: Resolution differences
3) What is the impact of Sense?
Table 3 Contrast to Noise Ratio
Scan 3 No sense Scan 4 Sense Cd 22 12.6 Pu 14.7 10.6 Cd Caudate, Pu Putamen
As expected sense lowers the CNR, yet if we look at the acquisition time
differences (No sense = 16 mins, sense = 6 mins) it quickly becomes evident that sense drastically reduces the scan time. When looking at the CNR to time
tradeoff we may want our sequence to include sense to limit acquisition time, regardless of the CNR loss. SNR calculations of these scans are in Appendix B.
4) What is the impact of field inhomogeneity (MPRAGE/MP2RAGE)?
Table 4 SNR of Left and Right Sphere
(scan 3)
Left Right
Cd 8.1 9
Pu 6.6 13.5
WM 3.8 3.7
Cd Caudate, Pu Putamen, WM White Matter
Briefly reiterating an MP2RAGE sequence is a T1-weighted scan, which is almost
identical to an MPRAGE, yet uses two images at different inversion time points to correct for field inhomogeneities (Marques et al., 2010). There is no direct scan comparison that can equivocally answer this question – What is the impact of
field inhomogeneity? – in our data, yet two pieces of evidence can help elucidate
an answer. Firstly, examining the SNR values of equidistant left and right sphere’s in a MPRAGE can indicate scan inhomogeneities. This rests on the premise that if there is no inhomogeneity, equidistant spheres of the same
structure in different hemispheres should have the same SNR. Table 4 shows the SNR values of left and right ROI’s in a MPRAGE (scan 3). Examining the results we can see there is not a large difference in signal when the spheres are
positioned close to each other (WM), yet as they are further apart (Pu) the inhomogeneity effects become apparent. The second piece of evidence adds explanation to why the further apart spheres have more variation in SNR. The image below (Figure 4) shows the consistent signal drop out in the right anterior part of the cerebellum on our MP2RAGE sequence. This signal drop out
becomes much more pronounced in a standard uncorrected MPRAGE
sequence. Since this signal dropout bleeds throughout the brain, it makes sense why a sphere in a ROI nearer to this dropout has less signal-to-noise than a further away sphere.
Figure 4: MP2RAGE cerebellum drop out
5) Are there differences between the MP2RAGE & MP2RAGME? (multiple echo’s) ?
Table 5 Contrast to Noise Ratio
Scan 6 MP2 0.7 Scan 7 MP4 0.7 Scan 8 Slab 0.6 Cd -10.14 -10.88 -6.60 Pu -9.04 -9.79 -5.50 MD -7.36 -6.22 -4.15
Table 5: MP2 0.7 MP2RAGE at 0.7mm isotropic, MP4 0.7 MP2RAGEME at 0.7mm isotropic, Slab 0.6 Partial slab scan using MP4RAGEME at 0.6mm isotropic.
As seen in Table 5, the only notable difference in contrast between the
MP2RAGE and the MP2RAGEME is within the thalamus, specifically the medial dorsal nucleus of the thalamus (MD).
We added a MP4RAGEME slab to get a higher resolution (0.6mm) in a similar scan time;; this is compared in table 5 to other sequences within the same individual to determine if the CNR is acceptable. This attempt to increase voxel resolution (0.6), while minimizing acquisition time resulted in a too noisy scan, as seen by the CNR calculations. SNR calculations of these scans are in Appendix B.
coronal fashion with an anterior – posterior direction, therefore when discussing the striatum this directionality in space will be used. The boundaries of the striatum are generally clear due to contrast changes on MPRAGE based sequences (MPRAGE, MP2RAGE & MP4RAGEME). Yet three main issues in defineing boundaries were identified pre-segmentation;; the Nucleus Accumbens (Ac) problem, Striatal cell bridges and the Islands or noise problem. A panel of researchers experienced in segmentation and anatomy discussed and reached the conclusions outlined in the guideline (Appendix C).
The Accumbens Nucleus (Ac) problem
The first issue, the ‘Ac problem’, involved how to define the borders of the Ac. Current in-vivo structural techniques are unable to distinguish the Ac from nearby structures, such as the bed nucleus (see Figure 5). A panel discussion resulted in the liberal approach of preferring to include the bed nucleus rather then exclude parts of the Ac. Therefore the segmentation cutoff was the lower tip of the lateral ventricle, as seen on Figure 5 of the MAI Atlas (Mai, Paxinos & Voss, 2008).
Figure 5: Lower tip of the lateral ventricle cut-off (Mai, Paxinos & Voss, 2008)
Striatal cell bridges problem
The ‘striatal cell bridges problem’ is located around the midpoint of the structure as the Cd and Pu split apart (as seen in Figure 6). This splitting process gives rise to multiple striatal cell bridges of varying sizes, the smaller bridges are undetectable on 7T MRI using MPRAGE, MP2RAGE and MP4RAGEME. This leads to the question, how to determine if a small cluster of pixels is noise or grey matter? The solution, agreed upon by a panel of experts, was to commence segmentation if the structure is a minimum of two pixels (voxel size = 0.7mm), with consideration to pervious slices and location. This cutoff was chosen in an attempt to include as much of the structure as possible. A more detailed report of these issues can be viewed in Appendix D.
small cluster of pixels is noise or grey matter? When segmenting in a coronal fashion, near the end of structure it will seem as if there are multiple Pu islands (Figure 7). A more accurate description would be to call them Pu peninsulas, as they are contacted to the structure as a whole. These structures vary in size and location through slices. The agreed upon solution, identical to the ‘striatal cell bridges problem’, was to segment if the structure is a minimum of two pixels (voxel size = 0.7mm), with consideration to pervious slices and location. This cutoff was chosen in an attempt to include as much of the peninsulas/islands as possible.
Figure 7: Putamen peninsulas (Mai, Paxinos & Voss, 2008)
Inter-rater Reliability (IRR)
The inter-rater reliability (IRR) was calculated via DICE and MHD,
aforementioned in the methods section. IRR calculations were necessary for two reasons. First and foremost, to determine segmentation similarity between the two independent segmenters. This IRR score could justify or invalidate the segmentation guideline. The second reason for calculating IRR was to use it as an indirect measure of sequence quality. This indirect measure would only work in the case of one type of sequence having systemically lower IRR scores.
Table 6 Inter-rater Reliability Measures
Dice MHD
MP2 Left 0.5mm (scan 1) 0.91 3.12*
MP2 Right 0.5mm (scan 1) 0.30 6.52*
MP2 Left (scan 2) 0.93 2.16
MP2 Right (scan 2) 0.94 1.84
TourMP2 Left (scan 5I) 0.93 1.72
TourMP2 Right (scan 5I) 0.94 1.62
TourMP2 Left (scan 6I) 0.94 1.76
TourMP2 Right (scan 6I) 0.93 1.86
MP4 Left (scan 7) 0.93 1.89
MP4 Right (scan 7) 0.92 1.93
MP4 Left (scan 9) 0.92 2.19
MP4 Right (scan 9) 0.93 2.09
MP2 MP2RAGE, TourMP2 MP2RAGE with a first inversion matched to Tourdias et al. 2014, MP4 synonymous
for a MP2RAGEME sequence
I these scans are identical yet with different subjects
* due to differing voxel sizes the MHD cannot be compared to the 0.7mm scans (all other scans listed)
The right mask of scan 1 was improperly saved by one of the segmenters, leading to a very low DICE score and a very high MHD (MHD=6.52). The DICE (range = 0.93-0.94) and MHD (range = 1.62-1.86) scores from scan’s 5 & 6 are very consistent, which is a promising result since they are identical sequences. This result seems to validate the segmentation guideline since the reliability across raters on the same sequence, yet with differing participants, is very consistent. The segmentation guidelines were developed with the purpose of facilitating consistency across segmenters;; therefore the across board high Dice scores and low MHD scores justify the use of the segmentation guidelines.
atlases, by eventually developing our own. This project was started via this report in a two-pronged manner. Firstly in when we develop an MR sequence with the ability to visualize multiple subcortical nuclei in differing modalities and secondly with the successful segmentation of the striatum. The first part, development of an MR sequence, was assisted through SNR/CNR calculations. Yet before segmenting the stratum guidelines had to be developed, only then could IRR calculations be completed. These IRR calculations offer a second utility to possibly identify sequences with consistently low scores, yet this result did not come to fruition. The discussion section follows the path of the project by firstly examining sequence development and then briefly discussing the segmentation of the striatum.
Sequence Development
The results from the SNR/CNR calculations helped confirm our conclusions derived by simply examining the scans. One particularly puzzling result is the inability to replicate Tourdias’ and colleagues (2014) MPRAGE sequence. This is especially odd since we also replicated their scan without sense, which resulted in higher CNR values, yet the intra-thalamic divisions were still not visible. Our finding is in stark contrast with their article. A myriad of factors could explain this failed replication such as, differing head coils, differing scanner manufactures and differing subjects. Following this lack of replication we determined three necessities for our MR sequence: 1) MP2RAGE 2) Resolution 3) Sense.
1) MP2RAGE
As seen on table 4, the inherent inhomogeneity effects from the 7T scanner were too large for the successful utilization of a MPRAGE sequence.
2) Resolution
Since increasing the voxel size did not yield any clear intra-thalamic divisions, we decided to revert back to 0.7mm isotropic. The two
sequences (scan 1 & 8) with voxel sizes bellow 0.7mm produced far too noisy scans for our purposes;; ergo we settled upon 0.7mm.
3) Sense
Not using sense increased the acquisition time almost threefold (from 6:30 to 16 mins) on a MPRAGE sequence at a relatively large voxel size
(1mm). We had already determined that correcting for field
inhomogeneities via an MP2RAGE sequence was necessary, along with preferring a smaller voxel size. Both of these factors lengthen acquisition time;; therefore we determined the use of sense was critical to keep the acquisition time reasonable.
The outcome of these decisions led to scans 5 & 6 (Table 1), which are standard MP2RAGE sequences at 0.7mm isotropic utilizing a first inversion (670ms) to match Tourdias and colleagues (2014). Since these scans were still unable to visualize intra-thalamic divisions our attention shifted towards QSM.
MP2RAGEME (MP4)
The most promising outcome of this project is the MP2RAGEME sequence, which is essentially a MPRAGE followed by four echoes. The first echo can be used as a second inversion, leading to the ability to compute a unified MP2RAGE scan. With the MP2RAGEME sequence it is also possible to compute a QSM map from the echoes. As aforementioned the QSM modality allows the
visualization of nuclei unattainable using MP2RAGE, such as the Gpi/e, RN, SN, STN (Keuken et al., 2014). Having the ability to derive both of these modalities from the same sequence allows multimodal segmentation without the need for registration. Scan 7 was the first implementation of the MP2RAGEME sequence. The CNR results (Table 5) confirmed that no contrast was lost in the striatum in comparison to the earlier MP2RAGE’s (scans 5 & 6).
Two conclusions resulted from this scanning session. Firstly, that we would stay at 0.7mm isotropic, rather then repeatedly attempting finer resolutions (scan 8). The second and more significant conclusion was to extend the last echo. We hypothesized that a later last echo would lead to increased T2* values, which in turn would benefit our QSM’s. Scan 9 was an attempt to match Keuken and colleagues (2014) last echo (29.57 ms). The last echo was heavily extended from 20.7 (scan 7) to 27.5 (scan 9), which in turn added six minutes more on the acquisition time. By chance this led to intra-thalamic divisions appearing. When attempting to lengthen the last echo, we had to slightly shorten the second
multiple modalities without the need for registration.
Figure 8: Intra-Thalamic Divisions (scan 9)
Segmentation of the Striatum
During segmentation no serious issues presented themselves. Only following segmentation was the improperly saved mask discovered. It is the
recommendation of the author to save the mask multiple times throughout the segmentation process and following completion reload the mask. One limitation of the design used was the lack of counterbalancing the order of scans between segmenters. This limitation is slightly mitigated by using trained segmenters on a limited amount of scans.
Inter-rater Reliability
Regardless of the sequence the DICE and MHD scores were fairly invariable. This consistency between segmenters can be interpreted as a success for the segmentation guideline. Yet on the flip slide it cannot offer any conclusions towards which scans are preferable to segment the Striatum. The MP2RAGEME sequence with a long last echo (scan 9) seems to have slightly larger MHD’s. This may be due to the border between the Cd and lateral ventricle being difficult to distinguish (Figure 8). Another possibility is time pressure upon the raters, since this scan was the last one to segment.
Conclusion
One of the more promising results from this project is the development of a MP2RAGEME sequence. This sequence has the potential for utility in the
segmentation of the striatum, intra-thalamic nuclei and a host of other subcortical nuclei. Further piloting will show if this result is replicable. The completed striatal segmentation guidelines (Appendix C) were successful in facilitating consistent segmenting. The one overarching limitation of this project is the lack of power behind the conclusions drawn in this report. Yet, this is an inherent limitation to the process of piloting as a whole. Further segmentation is needed to build a multimodal MR atlas of the subcortex from the MP2RAGEME sequence.
in neuroanatomy, 7, 40. Chicago
Beauchemin, M., Thomson, K. P. B., & Edwards, G. (1998). On the Hausdorff distance used for the evaluation of segmentation results. Canadian Journal
of Remote Sensing, 24(1), 3–8.
http://doi.org/10.1080/07038992.1998.10874685
Darki, F., Nemmi, F., Möller, A., Sitnikov, R., & Klingberg, T. (2016). Quantitative susceptibility mapping of striatum in children and adults, and its association with working memory performance. NeuroImage, 136, 208–214.
http://doi.org/10.1016/j.neuroimage.2016.04.065
de Hollander, G., Keuken, M. C., Bazin, P. L., Weiss, M., Neumann, J., Reimann, K., … Schäfer, A. (2014). A gradual increase of iron toward the medial- inferior tip of the subthalamic nucleus. Human Brain Mapping, 35(9), 4440– 4449. http://doi.org/10.1002/hbm.22485
Dice, L. R. (1945). Measures of the Amount of Ecologic Association Between Species Author ( s ): Lee R . Dice Published by : Wiley Stable URL : http://www.jstor.org/stable/1932409 Accessed : 08-04-2016 13 : 33 UTC Your use of the JSTOR archive indicates your acceptance of th. Ecology,
26(3), 297–302.
Dietrich, O., Raya, J. G., Reeder, S. B., Reiser, M. F., Schoenberg, S. O., Raya, G., … Schoenberg, S. O. (2007). Measurement of Signal-to-Noise Ratios in MR Images : Influence of Multichannel Coils, Parallel Imaging, and
Reconstruction Filters. Journal of Magnetic Resonance Imaging, 385(2), 375–385. http://doi.org/10.1002/jmri.20969
Hariz, M., Blomstedt, P., & Zrinzo, L. (2013). Future of brain stimulation: New targets, new indications, new technology. Movement Disorders, 28(13), 1784–1792. http://doi.org/10.1002/mds.25665
Huettel, SA., Song, AW., & McCarthy, G. Functional Magnetic Resonance
Imaging. 2nd ed. Sinauer associates Inc., Sunderland, MA: USA.
Keuken, M. C., Bazin, P. L., Crown, L., Hootsmans, J., Laufer, A., Müller-Axt, C., … Forstmann, B. U. (2014). Quantifying inter-individual anatomical variability in the subcortex using 7T structural MRI. NeuroImage, 94, 40–46.
http://doi.org/10.1016/j.neuroimage.2014.03.032
Kopell, B. H., & Greenberg, B. D. (2008). Anatomy and physiology of the basal ganglia: Implications for DBS in psychiatry. Neuroscience and Biobehavioral
Reviews, 32(3), 408–422. http://doi.org/10.1016/j.neubiorev.2007.07.004
Mai, J. K., Paxinos, G., & Voss, T. (2008) Atlas of the Human Brain (3rd ed.) New York, NY: Academic Press is an imprint of Elsevier.
Malone, D. A., Dougherty, D. D., Rezai, A. R., Carpenter, L. L., Friehs, G. M., Eskandar, E. N., … Greenberg, B. D. (2009). Deep Brain Stimulation of the Ventral Capsule/Ventral Striatum for Treatment-Resistant Depression.
Biological Psychiatry, 65(4), 267–275.
Marques, J. P., & Gruetter, R. (2013). New Developments and Applications of the MP2RAGE Sequence - Focusing the Contrast and High Spatial Resolution R1 Mapping. PLoS ONE, 8(7). http://doi.org/10.1371/journal.pone.0069294 Marques, J. P., Kober, T., Krueger, G., van der Zwaag, W., Van de Moortele, P.-
F., & Gruetter, R. (2010). MP2RAGE, a self bias-field corrected sequence for improved segmentation and T1-mapping at high field. NeuroImage, 49(2), 1271–1281. http://doi.org/10.1016/j.neuroimage.2009.10.002
Pruessmann, K. P., Weiger, M., Scheidegger, M. B., & Boesiger, P. (1999). SENSE: Sensitivity encoding for fast MRI. Magnetic Resonance in Medicine,
42(5), 952–962. http://doi.org/10.1002/(SICI)1522-
2594(199911)42:5<952::AID-MRM16>3.0.CO;;2-S
Saranathan, M., Tourdias, T., Bayram, E., Ghanouni, P., & Rutt, B. K. (2015). Optimization of white-matter-nulled magnetization prepared rapid gradient echo (MP-RAGE) imaging. Magnetic Resonance in Medicine, 73(5), 1786– 1794. http://doi.org/10.1002/mrm.25298
Schweser, F., Deistung, A., Sommer, K., & Reichenbach, J. R. (2013). Toward online reconstruction of quantitative susceptibility maps: Superfast dipole inversion. Magnetic Resonance in Medicine, 69(6), 1582–1594.
http://doi.org/10.1002/mrm.24405
Tourdias, T., Saranathan, M., Levesque, I. R., Su, J., & Rutt, B. K. (2014). Visualization of intra-thalamic nuclei with optimized white-matter-nulled MPRAGE at 7T. NeuroImage, 84, 534–545.
http://doi.org/10.1016/j.neuroimage.2013.08.069
Trépanier, L. L., Kumar, R., Lozano, a M., Lang, a E., & Saint-Cyr, J. a. (2000). Neuropsychological outcome of GPi pallidotomy and GPi or STN deep brain stimulation in Parkinson’s disease. Brain and Cognition, 42(3), 324–347. http://doi.org/10.1006/brcg.1999.1108
Udupa, K., & Chen, R. (2015). The mechanisms of action of deep brain
stimulation and ideas for the future development. Progress in Neurobiology,
133, 27–49. http://doi.org/10.1016/j.pneurobio.2015.08.001
Visser, E., Keuken, M. C., Douaud, G., Gaura, V., Bachoud-Levi, A. C., Remy, P., … Jenkinson, M. (2016). Automatic segmentation of the striatum and globus pallidus using MIST: Multimodal Image Segmentation Tool.
NeuroImage, 125, 479–497.
http://doi.org/10.1016/j.neuroimage.2015.10.013
Webb, A. G., & Van de Moortele, P. F. (2015). The technological future of 7 T MRI hardware. NMR in Biomedicine, (October 2014), n/a–n/a.
http://doi.org/10.1002/nbm.3315
Appendix Table of Contents
Appendix A
Sphere and true voxel sizes
p. 21
Appendix B
SNR results
p. 22
Appendix C
Striatum Guideline
p. 23 - 26
Appendix D
Three main guideline issues report
p. 27 - 34
Appendix A:
Sphere and true voxel sizes
Voxel size Num Voxels Volume Radius
0.92x1x0.92 139 116.42 3
0.64x0.7x0.64 137 39.35 2.1
0.58x0.58x0.6 147 29.92 1.9
0.47x0.5x0.47 123 13.85 1.5
All sizes are displayed in millimeter
Q1-2
Title: MP2RAGE SNR
Scan 1 Scan 2 Scan 5
Cd 5.1 5 3.5 Pu 3.9 6.2 3.7 WM 10.5 16.7 14 Q3 Title: MPRAGE SNR SNR No sense (scan 3) SNR Sense (scan 4) Cd 8.6 8.7 Pu 9.4 10.7 WM 3.8 2.2 Q5 Title: SNR MP2 v mp4 v MP4_slab Table 4: SNR Scan 6 MP2 0.7 Scan 7 MP4 0.7 Scan 8 Slab 0.6 Cd 3.5 13.4 6.9 Pu 5 15.8 7.8 WM 13.1 22 13.5 MD 5.5 19.5 9.5
Appendix C:
Striatum Guideline
Segmentation guidelines for MP2RAGE
Striatum:
The first segmentation includes the entire striatum outlined in continuous coronal slices with an anterior – posterior direction. Therefore segmentation commenced with the most anterior part of the caudate, inferiorly neighboring the lateral ventricle (when larger of 2x2 pixels). When segmenting use the Uni image along with the T1 to distinguish boundaries especially those neighboring the ventricles; this strategy is useful due to contrast changes.
It is important to 1) take into consideration pixel that go into noise or have strict boarders between slices and 2) when in doubt, utilize other dimensions.
Once the caudate and putamen are distinct, yet still connected structures, striatal cell bridges will connect the two. (As seen in image bellow)
To delineate the inferior boarder (Green arrows) extrapolate from the external capsule as it wraps around the putamen. End this boarder at the lower tip of the lateral ventricle, if difficult to discern switch between modalities and examine other fields of view. The motivation for this guideline is to be certain to include the Nucleus Accumbens, yet inevitably parts of the bed nucleus will also be included.
Secondly, for the superior boarder (Blue arrows) it is critical to exclude the Globus Pallidus (GP).
Lastly for the capsule internal boarder (Yellow arrow) end when the cell bridges have under a width of two pixels.
Start disconnecting when width is less than 2 pixels. (Taking contrast changes into consideration & previous slices)
Posterior Guidelines:
Unless clearly distinguishable ignore the tail of the caudate (TCd). Eventually the putamen forms islands, ignore these unless larger then 2x2 pixels. Take into consideration the pre & post slice and the location of the island when segmenting.
Appendix D:
Three main guideline issues report
Three main guideline issues & corresponding
decisions
1) Nucleus Accumbens (Ac)
Issue: Identifying the Nucleus Accumbens from other nearby subcortical structures (i.e.
Bed Nucleus)
Sequential, in intervals of two, coronal slices starting at the coordinates (207,192,217) and ending at (207,182,217) are displayed below (Image 1). The slices are displayed in an anterior to posterior direction; the changing axis, y, is listed on the slice.
Image 1 Coronal view of the Striatum
GP Globus Pallidus
This spatial series of slices highlight one of the critical aspects in the segmentation of the striatum, that is; where to place the cut-off of for the inclusion of the Nucleus Accumbens while excluding non-striatal related subcortical nuclei? Image two (207,186,217) shows potential cut-off lines for segmentation in a unified MP2RAGE modality.
To determine the optimal striatal cut-off line we consulted the Mai Atlas (2008) (Image 3). This resulted two distinct strategies regarding the Nucleus Accumbens (NAcc); (1) a
conservative approach (excluding large portions of the NAcc to ensure NAcc purity) or (2) a liberal approach (making sure to fully include the NAcc). Both of these approaches
have advantages and disadvantages.
Image 3 MAI Atlas
(Mai, Paxinos & Voss, 2008, p. 131,133)
The panel decided to have a cut-off at the lower tip of the lateral ventricle. This cut-off line can be viewed in the MAI atlas (Image 4) and in the previously foreseen unified MP2RAGE slice (207,186,217) (Image 5). A decision for the (2) liberal approach was chosen since we wanted our atlas to include the entirety of the NAcc. This approach will
inevitably include parts of the Bed Nucleus (BS), yet this was a necessary concession for the entire inclusion of the NAcc in our atlas.
Image 4 Striatal cut-off line on the MAI atlas
(Mai, Paxinos & Voss, 2008, p. 131,133)
which the Caudate and Putamen seem to split apart. In this process striatal cell bridges appear in varying sizes, as seen in the histological slice bellow of the MAI atlas (Image 6)
Image 6 Histological slice from the MAI atlas showing striatal cell bridges
(Mai, Paxinos & Voss, 2008, p. 128)
Below (Image 7) is presented a coronal anterior-posterior slice sequence on a MP2RAGE from Keuken et al. (2014) with an overlaid inter-rater segmentation mask. An inter-rater mask only displays the sections in which both segmenters agree upon.
Image 7 Inter-rater masks displayed on a T1-map MP2RAGE
Since it is very difficult to determine the difference between random noise and sub-millimetre cell bridges. Two methodological solutions were proposed; (1) ignore the cell
bridges or (2) to include them depending on their size.
The first proposal was initially contrived due to the lack of inter-rater consistency when segmenting striatal cell bridges. The images below show in coronal view the overlap of the strategy to ignore against the inter-rater masks; firstly side by side (Image 8) and then overlaid (Image 9).
Yellow mask (left) Inter-rater mask Red mask (right) An example of a mask ignoring striatal cell bridges
Image 9 Overlaid segmentation masks
Yellow mask Inter-rater mask Red mask overlaid An example of a mask ignoring striatal cell bridges
While this strategy seemed plausible in select slices, it quickly became evident this methodological strategy would not work for two reasons. Firstly, within subject’s larger striatal cell bridges show a consistent pattern through slices (as seen in image 10), secondly these large striatal cell bridges that show consistency across subjects also.
Image 10 Within subject consistency of striatal cell bridges
Red arrows highlighting large cell bridges show consistency through slices x(201) y(184-186) z(217)
For these reasons the panel decided to include a cell bridge with the minimum requirement of a width of two pixels. This criterion is valid only for segmentation in sequences with an isotropic voxel size of 0.7mm.
3) Islands or noise?
Issue: How to distinguish sub-millimetre Putamen islands from random noise.
The final guideline issue is that of the Putamen ‘islands’ which, in a coronal view, are located at the caudal part of the Striatum. As shown on the Mai atlas (Image 11) these ‘islands’ are of varying size, leading to a similar detection issue as evident in the striatal
cell bridges problem. These ‘islands’ have more peninsular characteristics, as they are
connected to the structure as a whole.
Image 12
(109,140,217)
Initially it may seem easy to identify the Striatal islands, as seen in Image 12 highlighted by red arrows. Yet upon closer inspection (Image 13) there are smaller clusters of pixels that cannot be easily identified as independent from noise (Yellow arrows).
Image 13 Islands or Noise?
(109,140,217)
Due to its location the two-pixel structure that arrow one is pointing towards on Image 13 could be assumed to be part of the Putamen. Yet not all cases are this easy, arrow two is much more difficult to identify as part of the Putamen due to its location.
The panel decided upon the same guidelines as with Striatal cell bridges; that is a Putamen island must have a minimum required width of two pixels with taking into consideration location and neighbouring slices. The rational behind taking neighbouring slices into consideration is due to the fact that these islands are essentially peninsulas, therefore should show across slice consistency.