Atlasing the Human Subcortex: Manual segmentation of the
Subthalamic Nucleus and Substantia Nigra at Ultra-High Resolution
Grace Pulsford
Student Number: 11120665
Supervisor: Martijn Mulder
Second Assessor: Anneke Alkemde
Research Institute: University of Amsterdam, Department of Integrative
Model Based Cognitive Neuroscience
MSc in Brain and Cognitive Sciences: Track Cognitive Neuroscience
Abstract
The human subcortex is made up of hundreds of individual and unique grey matter
nuclei, which combine together to form circuits, enabling various cognitive and motor
processes. Currently, however, in-vivo imaging at conventional field strengths of 3
Tesla (3T) or below is not sufficient to accurately visualise and delineate these small
structures. As a result, only 7% of subcortical nuclei are included in presently used
brain atlases, limiting research into the subcortex. The present study, therefore,
aimed to contribute towards the development of a more complete probabilistic atlas
of the human subcortex, by exploiting the increased signal to noise ratio (SNR) and
contrast to noise ratio (CNR) of 7T MRI. Two subcortical structures were focused on:
the subthalamic nucleus and substantia nigra. Individual segmentation protocols
were developed for both structures, which were subsequently followed to segment
the STN and SN on 9 ultra-high field quantitative susceptibility maps (QSM). Inter
and intra-rater reliability scores indicated that the STN segmentations were
consistent within and between raters, suggesting the protocol serves as reliable tool
for mapping the STN. Additionally, intra-rater reliability scores indicated the SN
segmentations were consistent between one rater. However, inter-rater reliability
measures are still needed in order to validate these findings.
Introduction
The human subcortex contains approximately 450 individual and unique grey matter structures. Overall, these make up approximately 25% of the entire volume of the brain (Forstmann et al., 2017). Combined, they form circuits that are involved in various cognitive and motor processes (Dunbar, 1992). Such processes are essential for everyday
functioning, and deficits to these structures are implicated in multiple neuropsychiatric and movement disorders (Rosenberg-Katz et al., 2016). Currently, however, in-vivo imaging at conventional field strengths of 3 Tesla (3T) or below is not sufficient to accurately visualise and delineate these small structures (de Hollander et al., 2015). As a result, only 7% of subcortical nuclei are included in presently used magnetic resonance imaging (MRI) brain atlases (Forstmann et al., 2017). Atlases provide a map of the brain, and define the spatial characteristics of different structures and regions. Importantly, atlases can be utilised in functional magnetic resonance (fMRI) studies, in order constrain the exact localisation of activity (Evans et al., 2012). A more complete subcortical atlas, is therefore vital for furthering our understanding of the function of specific subcortical structures. Recently, advances in MRI technology have led to the development of ultra-high field 7 tesla (7T) MRI. In comparison to lower field strengths, 7T MRI visualises many subcortical structures with an unprecedented amount of detail. As such, this study aimed to utilize the superior resolution of 7T MRI, in order to contribute towards the development of a more complete subcortical atlas. Two subcortical nuceli were focused on: the substantia nigra (SN) and subthalamic nucleus (STN).
Ultra-high field 7T MRI
Increasing the magnetic field strength of MRI to 7T greatly improves the spatial resolution available, producing images with increased signal to noise ratio (SNR) and contrast to noise ratio (CNR) (van der Kolk et al., 2013). This enables the visualisation of previously undetectable subcortical structures. While the detail of in-vivo MRI does not compare to that obtained via post-mortem data (Alkemade et al., 2012), the advantage of in-vivo data is that it can be collected from a range of many participants with different ages. Conversely, there are only a limited number of post-mortem brain specimens available. This is detrimental, as there is a huge amount of variability of the shape and size of subcortical structures between individuals (Keuken et al., 2014). Therefore, post-mortem work often misses this variability, whereas it can be accounted for using in-vivo MRI at 7T. As such, 7T MRI can be used to develop a probabilistic atlas – an atlas based on a number of brains, which provides a statistical confidence limit on positional variability of different structures (Mazziotta et al., 1995).
In order to develop a probabilistic atlas, segmentation of subcortical structures is needed. Segmenting a structure involves delineating it from the surrounding tissue, and defining its shape and location within the brain. This is carried out for the same structure over a number of different participants, in order to account for anatomical variability. The segmentation creates a mask of the structure, which is then registered to standard
stereotactic MNI space. Subsequently, all individual masks can be added together to create a probabilistic atlas (Keuken et al., 2014). Segmentation can be achieved either manually or automatically. Manual segmentation involves delineating the borders of the nuclei from neighbouring structures or tissue by hand, while automated segmentation uses algorithms to achieve this automatically. The disadvantage of manual segmentation is that it is highly labour intensive. Furthermore, the rater who is manually segmenting a structure can be susceptible to bias, basing their segmentation on how they believe the structure should look, rather than what is anatomically correct. However, current automated segmentation
procedures cannot yet fully account for the substantial amount of variability of the shape and location of many subcortical structures (Cabezas et al., 2011). Thus, manual segmentation is still considered the gold standard (Crum et al., 2006).
While manual segmentation of certain subcortical structures has already been completed (Keuken et al., 2014), there is still a need for a higher number of segmentations of the same structures, but from a wider range of ages. This is to create probabilistic atlases which account for the substantial morphological brain changes that occur over the adult lifespan (Peters, 2006). It is important that different raters follow a similar segmentation criterion for the same structure, to keep segmentation consistent. In order to control for this, segmentation protocols are needed. The protocol can be followed within and between different raters, in an attempt to keep reliability high. Moreover, as manual segmentation is a time consuming process, having comprehensive protocols available will help those who have not segmented the structure before move through it quicker.
In short, manual segmentation of subcortical structures using ultra-high resolution MR images is required in order to develop a more complete, probabilistic atlas of the human subcortex. In order to contribute towards this, the present study manually segmented two subcortical structures located within the basal ganglia (BG): the STN and SN.
The Subthalamic Nucleus & Substantia Nigra
The BG are situated within the subcortex, and consist of the striatum, internal and external segments of the globus pallidus (GPi, GPe), the STN and SN. This group of nuclei combine together to facilitate different essential cognitive and motor processes, including learning, memory, and reward reinforcement. The striatum receives information from a broad range of cortical areas, which is then sent either directly or indirectly through the GPe and STN, to the GPi and SN. In turn, these structures project to thalamic and brainstem nuclei in order to influence behaviour (Jin et al., 2014). While all structures within the BG play their own specific roles, vital for everyday functioning, this study focused on the SN and STN. Their specific locations within the BG and functions will be outlined below.
The Substantia Nigra
The SN is a curve linear shaped nucleus located inferior and medial the GP, and lateral to the mamillary body (Menke et al., 2010). It is made up of two anatomically and functionally distinct portions: the substantia nigra pars compacta (SNpc), and substantia nigra pars recticular (SNpr) (Pauli et al., 2015). GABAergic projections from the SNpr are part of the direct pathway within the BG, and are said to provide its primary output by inhibiting the thalamus. Conversely, dopaminergic neurons in the SNpc modulate synapses and excitability in the thalamus (Antal et al., 2014). Combined, the activity from the SNpc and SNpr is implicated in different functions including reward and learning (Zaghloul et al., 2009; Reynolds et al., 2001). Furthermore, the SN is part of the BG circuit which facilitates voluntary movement and inhibits competing movements (Venkateshappa et al., 2014). As such, damage to the SN is known to result in movement disorders such as Parkinson’s Disease (PD). Specifically, the progressive loss of dopaminergic neurons within the SNpc leads to an increase of spontaneous activity and periodic bursting in the STN (Remple et al., 2011). This causes many of the symptoms in PD, including tremor, rigidity and bradykinesia (Hirschmann et al., 2013; Tan et al., 2013).
The Subthalamic Nucleus
The STN is a lens shaped structure located medial to the inferior region of the GP, and lateral to the red nucelus, on the superiormedial edge of the SN. It is part of the striatal indirect pathway within the BG, receiving inhibitory input via the GPe (Fife et al., 2017). It is a is a densely populated nucleus, predominantly composed of glutamatergic excitatory projection neurons (Schwiezer et al., 2014). Projections from the STN are known to facilitate several processes including decision making, response control, and similar to the SN,
movement (Baunez, 2017; Obeso et al., 2014). As mentioned, it is the irregular activity of the STN which results in many of the symptoms of PD. In order to prevent this transmission of
pathologic bursting and oscillatory activity, deep brain stimulation (DBS) can be used. DBS for PD involves the placement of an electrode within the STN, which is subsequently stimulated. While the exact mechanisms of DBS are not yet known, one possible
explanation is that it dissociates input and output signals in the STN, disrupting abnormal information flow though the cortico-basal ganglia loop (Chiken & Nambu, 2016). It is currently the most effective treatment for PD, found to lessen motor fluctuations, reduce dyskinesia, and improve tremor, bradykinesia and rigidity (Limousin et al., 2008; Lagrange et al., 2002).
In addition to appropriate stimulation parameters, accuracy of electrode placement within the STN correlates with these motor improvements (Welter et al., 2014). Direct
targeting of the electrode within the STN can be achieved using 7T MRI (Ewert et al., 2017). Individual clinical scans are obtained, then by using landmarks, the location of the STN is approximated. However, due to small size of the STN, and its close proximity to the SN, targeting errors can occur. The mislocalisation of the electrode is considered to be the most common cause of adverse side effects and poor clinical outcome of DBS (Marks et al., 2009). It is therefore important to have more detailed information about the size and location of both structures, in order to prevent targeting errors. A high resolution atlas which includes DBS target structures provides this information, and thus may help to improve electrode placement in DBS surgery. The probabilistic nature of the atlas will be useful to take into account anatomical variability of DBS targets in different patients. This is another reason as to why manual segmentation of both the STN and SN is important.
Segmentation of the SN and STN
Previous studies have completed manual segmentation of both the STN and SN (Keuken et al., 2014; Du et al., 2011; Menke et al., 2009; Alkamde et al 2017). As the structures are located directly adjacent to one another, and are both small in size, it can be difficult to segment each nucleus accurately. In particular, defining the border between the two structures can be problematic, as the SN lies directly on the ventral lateral edge of the STN (Freestone et al., 2015). In the past, the contrast predominantly used to segment both structures has been T2*-weighted. This is because both nuclei are high in iron, and as iron is paramagnetic and decreases the transverse relaxation times, these nuclei appear as
hypointense on such images (Gelman et al., 1999). However, more recent research has found the novel MRI post-processing technique quantitative susceptibility mapping (QSM), to provide higher CNR in comparison to T2*-weighted images for the STN and SN (Liu et al., 2013; Deistung et al., 2013). QSM provides a quantitative assessment of the magnetic susceptibility of the tissue of interest, and is particularly beneficial when imaging regions high in iron content (Wisnieff et al., 2015). However, only a limited amount of studies have segmented these structures using QSM with the added benefit of 7T MRI. Furthermore, those that have done this mostly segment young, healthy participants, or those with PD. There is therefore a need for more segmentations based on a wider range of ages, without movement pathologies. The present study aimed to do this, using scans from age ranges between 18-80 years.
To summarise, the aim of this research was to produce reliable and consistent manual segmentations of the STN and SN using the added benefits of QSM MR images at ultra-high resolution. Individual segmentation protocols were developed for both structures. The STN protocol was strictly followed by two independent raters, who individually
segmented the same MR images. Subsequently, intra and inter-rater reliability scores were calculated. This was done to measure the reliability of the segmentations, and thus the reliability of the protocol. Due to time limitations, the SN was only segmented by one rater, following the protocol. The same rater segmented the same image twice, and intra-rater reliability scores were calculated to assess consistency between the two segmentation rounds. The protocols and masks created during segmentation are then available to contribute towards the probabilistic atlas.
Methods Participants
For the acquisition of the structural brain scans, 9 healthy participants (mean age 33.9) were randomly selected from a larger database of 72 scans from ages ranging between 18-80 years, acquired from the Integrative Model-based Cognitive Neuroscience Research Unit (IMCN) at the University of Amsterdam. All participants had no history of neurological, major medical or psychiatric disorders, and had normal to correct vision. The study was approved by the Ethics Review Board of the University of Amsterdam, and all participants gave their written and informed consent prior to scanning. Subjects were provided with monetary compensation for their participation.
Data acquisition of ultra-high resolution anatomical images
Structural images were acquired via 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. Whole brain images were acquired with an MP2RAGEME sequence. With this sequence, T1 and T2* images can both be collected all within the same scan. The signal of INV1 and INV2 can be used to fit a T1 function, of which a T1 map can be created, while the 4th echo is used to fit a T2* function to create a T2* map.
In addition, a quantitative susceptibility map (QSM) is calculated using the 4th echo of the
second inversion.
The MP2RAGEME consisted of 234 slices with an acquisition time of 17.84 mins (repetition time (TR) = 6298 ms, echo time (TE1/2/3/4) = 3/11.5/19/27.5 ms, inversion times TI1/TI2 = 670/ 2325.5 ms, flip angle = 4/4, voxel size = 0.7 mm. iHARPERELLA and iLSQR (as implemented in STIsuite V2.2; Li et al., 2011) were used to unwrap the phase data and remove the background phase for the QSM images (Li et al., 2015a, 2015b).
Manual segmentation of the STN and SN
Manual segmentation was performed using the FSL Viewer 5.0.9. To begin with, the STN and SN plus surrounding landmarks were localised on the MAI and DING atlases (Mai et al., 2015; Ding et al., 2016). Following this, T1 images, T2* images and QSM images of an independent training dataset of 8 scans were visually inspected, and used to locate the landmarks. Each contrast was visually compared in order to determine the optimal contrast to use for segmentation. This was based on the visibility of both the STN and SN. The comparison between T2* images and QSM images was particularly focused on, due to the fact recent studies have found an advantage of QSM over T2* (Liu et al., 2013; Deistung et al., 2013). Visual inspection displayed the advantage of QSM over the other contrasts. In particular, the difficult border between the STN and SN was clearer with this contrast in comparison to T2* (see Figure 1 for an example). QSM was therefore chosen as the contrast to segment both structures on.
Once the contrast was chosen, the right and left STN and SN of all scans within the training set were then segmented. The protocols for both the STN and SN segmentations were created based on these training segmentations, with reference to the atlases. See Appendix A for the complete STN and SN protocols.
Once the protocols were complete, the STN and SN were individually segmented strictly in accordance with the protocol guidelines. For every scan, separate masks were created for both the right and left hemisphere. Figure 2 gives an example of an STN and SN
segmentation in the coronal view. Both structures were segmented twice by the same rater. Subsequently, Dice Coefficient (Dice, 1945) and Modified Hausdorff distance (MHD)
(Dubuisson & Jain, 1994) scores were calculated as a measure of intra-rater reliability between the two segmentation rounds. The Dice score measures the degree to which two different volumes are associated with one another. A coefficient of association of 1.0 shows that the two volumes occur together in exactly the number of sample units expected by chance. Any coefficient smaller than 1.0 shows that they occur together in fewer samples. The MHD describes the similarity between two objects. The advantage of this calculation is ‘its value increases monotonically as the amount of difference between the two sets of edge points increases’ (Dubuisson & Jain, 1994), thus it is not biased by the structure you have segmented increasing in volume. Two tests were carried out rather than one to have an extra measure of reliability. Following this, the STN was segmented by a second,
independent rater. Again, Dice Coefficient and MHD scores were calculated. This was used as a measure of inter-rater reliability between the two raters. The SN was not segmented by a second independent rater due to time limitations.
Figure 1. Comparison between QSM (left) and T2*-weighted (right) images. QSM = Quantitative susceptibility mapping, STN = subthalamic nucleus, SN substantia nigra.
Figure 2. QSM image of bilateral STN and SN in the coronal view. Left side shows structures not segmented. Right side shows example of STN (red) and SN (blue) segmentation.QSM = Quantitative susceptibility mapping, STN = subthalamic nucleus, SN substantia nigra.
Results
The STN and SN were segmented and the resulting volumes were calculated. The mean(std) volume of the STN was 154.44 mm3 (30.88), while the mean volume of the SN
was 571.93 mm3 (73.32) (see Table 1 for all volumes). These volumes are in the range of
previous studies which have segmented these structures (Alkemade et al., 2017; Menke et al., 2010).
Dice scores and MHD were then calculated. The mean(std) intra-rater Dice score was 0.91(.05) for the STN, and 0.95 (0.03) for the SN. The Mean (std) MHD was 0.13 (.05) for the STN, and 0.23 (0.05) for the SN (see table 2 for all intra-rater results). The mean(std) inter-rater reliability Dice score for the STN was 0.77 (0.08), while the mean MHD was 0.20(0.04). Overall the intra and inter-rater reliability scores indicate excellent agreement within and between raters, in line with previous studies (Keuken et al., 2014).
Table 1.
Combined volumes (mm3) of 1st and 2nd segmenting sessions of first rater for the SN and STN, plus STN volumes of second rater. SN = substantia nigra, STN = subthalamic nucleus, LH = left hemisphere, RH = right hemisphere.
Scan number SN STN first rater STN second rater
LH RH LH RH LH RH 3 676.91 672.74 181.20 149.85 144.38 129.43 8 508.65 477.16 149.56 142.80 80.82 93.19 9 578.11 580.27 123.25 147.84 100.38 161.98 10 560.28 528.78 115.34 103.69 72.19 107.28 11 560.28 584.73 148.84 137.05 113.61 133.45 12 550.36 443.22 155.31 140.65 100.38 103.54 17 476.01 500.89 131.01 192.13 76.22 106.13 18 666.84 635.20 204.50 164.66 149.56 127.13 19 661.81 632.62 213.85 175.30 120.80 136.04 Mean (std) 582.14 (70.13) 561.73 (75.01) 158.45 (35.43) 150.44 (24.89) 106.48 (26.61) 122.01 (20.12)
Table 2.
Intra-rater reliability results of each individual scan, per structure, per hemisphere. MHD = Modified Hausdorff distance,SN = substantia nigra, STN = subthalamic nucleus, LH = left hemisphere, RH = right hemisphere.
Table 3.
Inter-rater reliability results of the STN for each indivudal scan, per hemisphere. MHD = Modified Hausdorff distance, STN = subthalamic nucleus, LH = left hemisphere, RH = right hemisphere. Dice MHD Scan number SN STN SN STN LH RH LH RH LH RH LH RH 3 0.96 0.96 0.85 0.93 0.24 0.25 0.20 0.11 8 0.94 0.96 0.96 0.98 0.24 0.19 0.19 0.03 9 0.97 0.95 0.81 0.83 0.19 0.22 0.19 0.18 10 0.96 0.96 0.96 0.91 0.20 0.19 0.08 0.09 11 0.93 0.94 0.92 0.95 0.27 0.26 0.13 0.09 12 0.98 0.96 0.92 0.92 0.13 0.18 0.15 0.13 17 0.87 0.89 0.88 0.87 0.34 0.31 0.15 0.20 18 0.95 0.96 0.94 0.93 0.23 0.22 0.14 0.12 19 0.95 0.92 0.90 0.95 0.26 0.28 0.19 0.11 Mean (std) 0.95 (0.03) 0.95 (0.02) 0.90 (0.04) 0.92 (0.04) 0.23 (0.05) 0.24 (0.04) 0.14 (0.04) 0.12 (0.05) Dice MHD Scan Number LH RH LH RH 3 0.87 0.82 0.16 0.19 8 0.71 0.75 0.21 0.20 9 0.78 0.86 0.18 0.17 10 0.70 0.81 0.19 0.15 11 0.86 0.89 0.15 0.15 12 0.75 0.80 0.23 0.20 17 0.58 0.64 0.24 0.27 18 0.81 0.77 0.22 0.20 19 0.73 0.78 0.26 0.23 Mean (std) 0.75(0.08) 0.79(0.07) 0.21(0.03) 0.19(0.04)
Discussion
The development of ultra-high field 7T MRI has enabled the visualisation and delineation of many small subcortical structures, previously unable to be seen at lower field strengths (Strotmann et al., 2014). The present study aimed to take advantage of the improved resolution of 7T images in order to manually segment the STN and SN. By doing so, the project contributed to the development of a probabilistic atlas of the subcortex. First, segmentation protocols were individually created for each structure. The STN protocol was followed by two independent raters. Both intra and inter-rater reliability tests showed segmentations were consistent. The SN protocol was followed by one rater, and intra-rater results indicated segmentations between the same rater were reliable. The implications of these results will be discussed below.
Segmentation
Reliability of segmentations
The high intra and inter-rater reliability scores for the STN suggest that the
segmentation protocol was followed accurately by both raters. As such, this signifies that the segmentations created are reliable enough to contribute towards a probabilistic atlas. A probabilistic atlas including the STN can be utilised in fMRI studies, in order to precisely localise activity. As such, more research into the specific functions of the STN can be carried out. Furthermore, the high agreement between the two independent raters indicates that this protocol is a reliable tool, which can be followed by other raters to create additional,
consistent manual segmentations of the STN. This is important, as more information can be gathered about inter-individual structural differences of the STN. This may be relevant for DBS surgery, as knowing more of the variability of the shape and location of the STN may help to achieve more accurate electrode targeting in PD. In addition, the high intra-rater reliability sores for the SN give a preliminary indication that the protocol can be used by other raters to create consistent manual segmentations. However, inter-rater reliability must be measured before this can be confirmed.
Problems with segmentation
While the reliability scores of the segmentations were high, there were specific problems which arose when segmenting the STN and SN. These are important to note, as such issues most likely contributed to reducing the intra and inter-rater reliability scores. The problems will be outlined below.
Border between the STN and SN
The border between the STN and SN is notoriously difficult to define. In order to determine this boundary as accurately as possible, QSM images were chosen. While the advantages of QSM were not quantitatively tested in the present study, visual comparison showed this border was much clearer on QSM images in comparison to the other contrasts. This is in-line with recent research finding improved CNR of QSM when imaging the STN and SN (Liu et al., 2013; Deistung et al., 2013). However, in MRI, even with a sophisticated contrast mechanism such as QSM, and the added benefit of ultra-high resolution, the voxel size is significant in comparison to the scale of the object imaged. Therefore, at a border where two different structures meet, the signal intensity within a single voxel can be a combination of the contributions of both tissues, rather than a single tissue type (Ballester et al., 2002). This is known as partial volume effects. These partial volume effects impacted the border between the STN and SN, as the SN lies directly on the ventral lateral edge of the
STN (see figure 3 for example). This was the most prominent problem throughout
segmentation of both structures, exaggerated by the fact that both are high in iron, and thus simultaneously show as hyperintense on QSM images (Barbosa et al., 2015). The difference in signal intensity between them was therefore small in comparison to the surrounding tissue. In order to overcome this issue, voxels were individually selected and inspected the coronal, axial and sagittal views. By using information from all viewpoints, a more accurate decision could be made of where the border should be. If an informed decision could not be made by using this technique, a priori knowledge of the location of the border was relied upon using information from the MAI and Ding atlases (Mai et al., 2015; Ding et al., 2016), and knowledge from previous segmentations. This could make definition of this border quite arbitrary.
Motion Artefacts
The amount of anatomical detail available at 7T is greatly improved in comparison to lower magnetic field strengths. In the present study, voxel resolution was 0.7mm3, permitting
fairly precise segmentation. However, by increasing the resolution, image quality is also increasingly sensitive to small involuntary motion of the subject (Gallichan et al., 2016). This creates motion artefacts, which systematically distort MRI data, even at a sub-millimetre level (Dosenbach., 2017). If such distortions impact an image, accurate segmentation becomes difficult, particularly for smaller structures like the STN and SN. Due to their smaller volume, slight movement has a large impact on boundary definition. Figure 4 shows an example of an image of the STN and SN most likely effected by motion, the blue arrow indicates approximately where the border between the two structures should be. In a case such as this, again a priori knowledge of the shape and location of both structures was relied upon, combined with knowledge from previous segmentations. This may have produced biased masks to what the rater thought the borders should look like, rather than what was anatomically correct.
An alternative explanation for the blurring in some of the scans is that there may have been a problem with either the online or offline processing of the QSM. A way to check for this is to examine different scans of the same subject. If it is only the QSM image which seems to be distorted, it is most likely a case of disrupted QSM processing. In such cases, the QSM scan should be discarded.
Fig 3. QSM image of the STN and SN impacted by partial volume effects. Black arrow indicates where the border between the two structures is. It is difficult to assign voxels on this border. QSM = Quantitative susceptibility mapping, STN = subthalamic nucleus, SN substantia nigra.
.
Fibre tracts between the STN & GP and SN & GP
The STN and SN are part of an intricate, interconnected network within the BG. As such, fibre tracts connect both nuclei to neighbouring structures, facilitating many different cognitive and movement processes (Pujol et al., 2017; Ziegler et al., 2014). These fibre tracts can also be visualized with 7T MRI, appearing as hyperintense on QSM images (Langkammer et al., 2013). However, as both fibre tracts and the structures themselves have similar levels of intensity, it is sometimes difficult to separate the two. Figure 5 gives an example of this. The blue arrow points towards the lateral tip of the SN, which appears to blur into the globus pallidus internal part (GPi). This blurring is most likely fibre tracts.
Furthermore, it is also possible to see what looks like an extension of the STN on its superior edge, displayed by the green arrow. This is also most likely fibre tracts, and should not be included in the STN mask. When segmenting, in order to determine where the structure finishes and the fibre tract begins, again a combination of all three views was used. By combining the information from each view-point, a more informed decision could be made of where to end segmentation.
Fig 4. QSM image affected by motion artefacts. Blue arrow indicates
approximately where the border between the STN and SN is. QSM = Quantitative susceptibility mapping, STN = subthalamic nucleus, SN substantia nigra.
Fig 5. QSM image displaying fiber tracts from the SN and STN. Blue arrow indicates fiber tracts from the SN to the GPi, green arrow displays fiber tracts from the STN. QSM = Quantitative susceptibility mapping, STN = subthalamic nucleus, SN substantia nigra.
Manual vs automated segmentation
The present study chose to manually segment the STN and SN as currently it is considered the gold standard (Crum et al., 2006). However, the specific problems outlined above exemplify how manual segmentation can be susceptible bias. The rater must make informed decisions by combining different sources of information. However, this can lead to mistakes, or a rater segmenting the structure the way they believe it should be, rather than what is anatomically correct. Another limitation of manual segmentation is that it is an intensive and time consuming task. In order to develop a probabilistic atlas, many
segmentations of the same structure are needed to account for variability. The number that is required depends on the group which the atlas is indented to represent. For instance, there may be less variation within a relatively homogenous group such as young and healthy individuals, in comparison to a group including a wide range of ages, or those with
pathological conditions. The STN, for example, shows a lateral shift with age, possibly due to loss of tissue within the internal capsule, which is located on its lateral border (Pereira et al., 2016). This variation therefore needs to be accommodated for with a higher amount of segmentations. Nevertheless, even in homogenous groups, anatomical variability is still very high (Massey et al., 2012). Thus, a significant amount of time is needed in order to manually trace many MR images, for any probabilistic atlas.
The development of more sophisticated and accurate automated segmentation procedures is therefore desirable, to quicken the segmentation process. Automated
algorithms that are based on statistical shape and appearance models integrate information about the mean and variance of structures shape and intensity on MRI scans. Currently, however, there is not enough information in regards to the variability of the shape and location of subcortical structures - including the STN and SN - for precise automatic segmentation. The anatomical information gathered in the present study can therefore be used to contribute towards the development of more advanced automated segmentation processes. Once developed, the creation of probabilistic subcortical atlases will be greatly sped up (Fortsmann et al., 2017).
Limitations
An important limitation of this study is the lack of inter-rater reliability tests for the SN segmentations. To begin with, the SN protocol was created by the one rater who segmented the SN. While this protocol was developed using an independent set of scans to the ones tested on, it is likely that this rater had a bias from creating the protocol. Moreover, as the same rater then segmented the same scan twice, they most likely developed a memory bias from the first segmentation to the next. One control put in place to prevent this was the randomising of the order of segmentations, in that a different, random order was used in the first and second segmenting rounds. However, while this may have prevented some memory bias, it is still possible the rater remembered how they segmented the image the first time round. It is therefore not yet possible to claim that the SN protocol will create reliable masks between raters. In the future, the SN protocol should be followed by another independent rater, and inter-rater reliability scores should be calculated in order to validate the current findings. Only then can the protocol and current segmentations be defined as reliable.
A second limitation is the limited number of segmentations completed. Overall, only 9 STN and SN masks were created. This was taken from a database with ages ranging from 18-80. While this is a preliminary step into mapping the morphological brain changes which occur over the adult life-span, many more segmentations are needed to account for the large intra- and inter-individual variability of the shape and location of both structures (Keuken et al., 2014). However, the STN protocol, and once validated, the SN protocol, are available for others to follow in order to produce more segmentations.
A further, more general limitation, is that while increase of magnetic field strength to 7T dramatically improves the detail of small subcortical structures in comparison to 3T, it does not compare to the level of detail obtained via post mortem myelo- and
cytoarchitectonical work. Dissection and segmentation of post mortem brains can result in a resolution as high as 6um, providing superb anatomical detail in comparison to MRI
(Alkemade et al., 2012). However, as discussed there is limited amount of post-mortem data in comparison to in-vivo images. Thus, in order to validate and extend the current findings, the in-vivo results should be compared with additional histochemical data. Finally, while the use of QSM in the present study, and in many other studies (Liu et al., 2013; Deistung et al., 2013), is an exciting approach for the visualisation and quantification of the location, shape and morphometric changes of nuclei such as the STN and SN, there is still room to further to optimize scan parameters for the visualisation of subcortical structures. In vivo-MRI is
continuously adapting and exciting developments are gradually occurring. However, further progress can still be made to provide even more detailed images of the subcortex and the structures within it.
Conclusions
Overall, the STN was segmented with both high intra and inter-rater reliability, indicating the segmentations are reliable enough to contribute towards a probabilistic atlas, and that the protocol can be used to create additional consistent segmentations. Additionally, the SN was segmented with high intra-rater reliability. This gives a preliminary indication that both the SN segmentations and protocol are reliable. However, inter-rater reliability must be
measured in order to confirm this. Finally, the information gathered from the present study in regards to the anatomical variability of both structures, can be used to develop more
sophisticated automated segmentation algorithms. Once developed, segmentation
processes will become much quicker, speeding up the creation of a probabilistic atlas of the subcortex.
Acknowledgments
I would firstly like to thank my two supervisors Martijn Mulder and Anneke Alkemade for helping me throughout this project, especially their support and patience. I would also like to thank Max Keuken for giving up his time to thoroughly answer any questions I had – big or small. I would also like to thank Birte and the rest of the Fortsmann lab, for making this internship a great experience.
References
Alkemade, A., De Hollander, G., Kitchen, MC, Schäfer, A., Ott, DV, Schwarz, J., ... & Forstmann, BU (2017). Comparison of T2 * -weighted and QSM contrasts in Parkinson's disease to visualize the STN with MRI. PloS one ,
Alkemade, A., Yi, CX, Pei, L., Harakalova, M., Swaab, DF, la Fleur, SE, ... & Kalsbeek, A.
(2012). AgRP and NPY expression in the human hypothalamic infundibular nucleus correlate with body mass index, whereas changes in αMSH are related to type 2 diabetes. The Journal of Clinical Endocrinology & Metabolism ,
Ballester, M. A. G., Zisserman, A. P., & Brady, M. (2002). Estimation of the partial volume effect in MRI. Medical Image Analysis, 6(4), 389-405
Barbosa, J. H. O., Santos, A. C., Tumas, V., Liu, M., Zheng, W., Haacke, E. M., & Salmon, C. E. G. (2015). Quantifying brain iron deposition in patients with Parkinson's disease using quantitative susceptibility mapping, R2 and R2. Magnetic resonance imaging, 33(5), 559-565.
Baunez, C. (2017). 45-The Subthalamic Nucleus at the Nexus of Decision-Making Processes. Biological Psychiatry, 81(10), S19.
Cabezas, M., Oliver, A., Lladó, X., Freixenet, J., & Cuadra, M. B. (2011). A review of atlas-based segmentation for magnetic resonance brain images. Computer methods and programs in
biomedicine, 104(3), e158-e177.
Chiken, S., & Nambu, A. (2016). Mechanism of deep brain stimulation: inhibition, excitation, or disruption?. The Neuroscientist, 22(3), 313-322.
Chu, R., Hurwitz, S., Tauhid, S., & Bakshi, R. (2017). Automated segmentation of cerebral deep gray matter from MRI scans: effect of field strength on sensitivity and reliability. BMC neurology, 17(1), 172.
Crum, WR, Camara, O., & Hill, DL (2006). Generalized overlap measures for evaluation and validation in medical image analysis. IEEE Transactions on Medical Imaging, 25 (11), 1451-1461.
Deistung, A., Schäfer, A., Schweser, F., Biedermann, U., Turner, R., & Reichenbach, J. R. (2013). Toward in vivo histology: a comparison of quantitative susceptibility mapping (QSM) with magnitude-, phase-, and R 2⁎-imaging at ultra-high magnetic field strength. Neuroimage, 65, 299-314.
Despotović, I., Goossens, B., & Philips, W. (2015). MRI segmentation of the human brain: challenges, methods, and applications. Computational and mathematical methods in medicine, 2015.
Dice, L. R. (1945). Measures of the amount of ecologic association between species. Ecology, 26(3), 297-302
Dunbar, R. I. (1992). Neocortex size as a constraint on group size in primates. Journal of human evolution, 22(6), 469-493.
Ding, S. L., Royall, J. J., Sunkin, S. M., Ng, L., Facer, B. A., Lesnar, P., ... & Stevens, A. (2016). Comprehensive cellular‐resolution atlas of the adult human brain. Journal of Comparative Neurology, 524(16), 3127-3481.
Du, G., Lewis, M. M., Styner, M., Shaffer, M. L., Sen, S., Yang, Q. X., & Huang, X. (2011). Combined R2* and diffusion tensor imaging changes in the substantia nigra in Parkinson's disease. Movement Disorders, 26(9), 1627-1632.
Dubuisson, M. P., & Jain, A. K. (1994, October). A modified Hausdorff distance for object matching. In Pattern Recognition, 1994. Vol. 1-Conference A: Computer Vision & Image Processing.,
Proceedings of the 12th IAPR International Conference on (Vol. 1, pp. 566-568). IEEE.
Dosenbach, N. U., Koller, J. M., Earl, E. A., Miranda-Dominguez, O., Klein, R. L., Van, A. N., ... & Wesevich, V. (2017). Real-time motion analytics during brain MRI improve data quality and reduce costs. NeuroImage.
Evans, A. C., Janke, A. L., Collins, D. L., & Baillet, S. (2012). Brain templates and atlases. Neuroimage, 62(2), 911-922.
Ewert, S., Plettig, P., Li, N., Chakravarty, M., Collins, L., Herrington, T., ... & Horn, A. (2017). Toward defining deep brain stimulation targets in MNI space: A subcortical atlas based on multimodal MRI, histology and structural connectivity. NeuroImage.
Fife, K. H., Gutierrez-Reed, N. A., Zell, V., Bailly, J., Lewis, C. M., Aron, A. R., & Hnasko, T. S. (2017). Causal role for the subthalamic nucleus in interrupting behavior. eLife, 6.
Forstmann, B. U., de Hollander, G., van Maanen, L., Alkemade, A., & Keuken, M. C. (2017). Towards a mechanistic understanding of the human subcortex. Nature Reviews Neuroscience, 18(1), 57-65.
Freestone, P. S., Wu, X. H., de Guzman, G., & Lipski, J. (2015). Excitatory drive from the subthalamic nucleus attenuates GABAergic transmission in the substantia nigra pars compacta via
endocannabinoids. European journal of pharmacology, 767, 144-151.
Gelman, N., Gorell, J. M., Barker, P. B., Savage, R. M., Spickler, E. M., Windham, J. P., & Knight, R. A. (1999). MR imaging of human brain at 3.0 T: preliminary report on transverse relaxation rates and relation to estimated iron content. Radiology, 210(3), 759-767.
Jin, X., Tecuapetla, F., & Costa, R. M. (2014). Basal ganglia subcircuits distinctively encode the parsing and concatenation of action sequences. Nature neuroscience, 17(3), 423-430.
.
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.
Lagrange, E., Krack, P., Moro, E., Ardouin, C., Van Blercom, N., Chabardes, S., ... & Pollak, P. (2002). Bilateral subthalamic nucleus stimulation improves health-related quality of life in PD. Neurology, 59(12), 1976-1978.
Langkammer, C., Liu, T., Khalil, M., Enzinger, C., Jehna, M., Fuchs, S., ... & Ropele, S. (2013). Quantitative susceptibility mapping in multiple sclerosis. Radiology, 267(2), 551-559.
Li, W., Wang, N., Yu, F., Han, H., Cao, W., Romero, R., ... & Liu, C. (2015). A method for estimating and removing streaking artifacts in quantitative susceptibility mapping. Neuroimage, 108, 111-122.
Limousin, P., & Martinez-Torres, I. (2008). Deep brain stimulation for Parkinson’s disease. Neurotherapeutics, 5(2), 309-319.
Liu, T., Eskreis-Winkler, S., Schweitzer, A. D., Chen, W., Kaplitt, M. G., Tsiouris, A. J., & Wang, Y. (2013). Improved subthalamic nucleus depiction with quantitative susceptibility
mapping. Radiology, 269(1), 216-223.
Mai, J. K., Majtanik, M., & Paxinos, G. (2015). Atlas of the human brain. Academic Press.
Marks, W. A., Honeycutt, J., Acosta, F., & Reed, M. (2009, June). Deep brain stimulation for pediatric movement disorders. In Seminars in pediatric neurology (Vol. 16, No. 2, pp. 90-98). WB Saunders.
Massey, LA, Miranda, MA, Zrinzo, L., Al-Helli, O., Parkes, HG, Thornton, JS, ... & Holton, JL (2012). High resolution MR anatomy of the subthalamic nucleus: imaging at 9.4 T with histological validation. Neuroimage ,
Menke, R. A., Jbabdi, S., Miller, K. L., Matthews, P. M., & Zarei, M. (2010). Connectivity-based segmentation of the substantia nigra in human and its implications in Parkinson's
disease. Neuroimage, 52(4), 1175-1180.
Menke, R. A., Scholz, J., Miller, K. L., Deoni, S., Jbabdi, S., Matthews, P. M., & Zarei, M. (2009). MRI characteristics of the substantia nigra in Parkinson's disease: a combined quantitative T1 and DTI study. Neuroimage, 47(2), 435-441.
Number, M., Beneduce, BM, & Regehr, WG (2014). The substantia nigra conveys target-dependent excitatory and inhibitory outputs from the basal ganglia to the thalamus. Journal of
Neuroscience , 34 (23), 8032-8042.
Obeso, I., Wilkinson, L., Casabona, E., Speekenbrink, M., Luisa Bringas, M., Álvarez, M., ... & Obeso, J. A. (2014). The subthalamic nucleus and inhibitory control: impact of subthalamotomy in Parkinson’s disease. Brain, 137(5), 1470-1480.
Pauli, W. M., Larsen, T., Collette, S., Tyszka, J. M., Seymour, B., & O'Doherty, J. P. (2015). Distinct contributions of ventromedial and dorsolateral subregions of the human substantia nigra to appetitive and aversive learning. Journal of Neuroscience, 35(42), 14220-14233.
Pereira, JL, Furie, S., Sharim, J., Yazdi, D., DeSalles, AA, & Pouratian, N. (2016). Lateralization of the subthalamic nucleus with age in Parkinson's disease. Basal ganglia ,
Peters, R. (2006). Ageing and the brain. Postgraduate medical journal, 82(964), 84-88.
Pujol, S., Cabeen, R., Sébille, S. B., Yelnik, J., François, C., Fernandez Vidal, S., ... & Kikinis, R. (2017). In vivo exploration of the connectivity between the subthalamic nucleus and the globus pallidus in the human brain using multi-fiber tractography. Frontiers in neuroanatomy, 10, 119.
Remple, M. S., Bradenham, C. H., Kao, C. C., Charles, P. D., Neimat, J. S., & Konrad, P. E. (2011). Subthalamic nucleus neuronal firing rate increases with Parkinson's disease progression. Movement Disorders, 26(9), 1657-1662.
Reynolds, J. N., Hyland, B. I., & Wickens, J. R. (2001). A cellular mechanism of reward-related learning. Nature, 413(6851), 67-70.
Rosenberg-Katz, K., Herman, T., Jacob, Y., Kliper, E., Giladi, N., & Hausdorff, J. M. (2016). Subcortical Volumes Differ in Parkinson’s Disease Motor Subtypes: New Insights into the Pathophysiology of Disparate Symptoms. Frontiers in human neuroscience, 10.
Schweizer, N., Pupe, S., Arvidsson, E., Nordenankar, K., Smith-Anttila, C. J., Mahmoudi, S., ... & Leão, R. N. (2014). Limiting glutamate transmission in a Vglut2-expressing subpopulation of the subthalamic nucleus is sufficient to cause hyperlocomotion. Proceedings of the National Academy of Sciences, 111(21), 7837-7842.
Strotmann, B., Heidemann, R. M., Anwander, A., Weiss, M., Trampel, R., Villringer, A., & Turner, R. (2014). High‐resolution MRI and diffusion‐weighted imaging of the human habenula at 7 tesla. Journal of Magnetic Resonance Imaging, 39(4), 1018-1026.
Tan, H., Pogosyan, A., Anzak, A., Foltynie, T., Limousin, P., Zrinzo, L., ... & Brown, P. (2013). Frequency specific activity in subthalamic nucleus correlates with hand bradykinesia in Parkinson's disease. Experimental neurology, 240, 122-129.
van der Kolk, A. G., Hendrikse, J., Zwanenburg, J. J., Visser, F., & Luijten, P. R. (2013). Clinical applications of 7T MRI in the brain. European journal of radiology, 82(5), 708-718.
Venkateshappa, C., Harish, G., Mythri, R., Mahadevan, A., Shankar, S. K., & Bharath, S. M. (2014). Increased oxidative damage and decreased antioxidant function in aging human substantia nigra compared to striatum: Implications for Parkinson's disease. Movement Disorders, 29, 35.
Wisnieff, C., Ramanan, S., Olesik, J., Gauthier, S., Wang, Y., & Pitt, D. (2015). Quantitative susceptibility mapping (QSM) of white matter multiple sclerosis lesions: interpreting positive
Welter, M. L., Schüpbach, M., Czernecki, V., Karachi, C., Fernandez-Vidal, S., Golmard, J. L., ... & Mesnage, V. (2014). Optimal target localization for subthalamic stimulation in patients with Parkinson disease. Neurology, 82(15), 1352-1361.
Zaghloul, K. A., Blanco, J. A., Weidemann, C. T., McGill, K., Jaggi, J. L., Baltuch, G. H., & Kahana, M. J. (2009). Human substantia nigra neurons encode unexpected financial
rewards. Science, 323(5920), 1496-1499.
Ziegler, E., Rouillard, M., André, E., Coolen, T., Stender, J., Balteau, E., ... & Garraux, G. (2014). Mapping track density changes in nigrostriatal and extranigral pathways in Parkinson's
Appendix A
Subthalamic Nucleus Segmentation Protocol
Note this protocol will focus on the left subthalamic nucleus (STN). The same segmentation procedure applies to the right STN. Segment the left and right STN separately, creating individual masks for each.
1.0 Start up fsl
• Open fslview
• MATE Terminal • Type ‘fslview’ • Open the QSM image
• File > open > home/public/HumanAtlas/subcortex/R01/
subcortex_scannumber_subjectnumer_R01_INV2_e4_P_QSM)
• Adjust the contrast so that you can see the definition between different structures; around -0.3/0.3 is a good guide. However, what is optimal may vary between scans.
• Create a mask
• File > create mask 2.0 Refer to the MAI atlas
• Use the following information, combined with figure 1, to become familiarized with the location of the STN and structures surrounding it. All structures mentioned will appear as hyperintense on QSM images.
• The STN is located medial to the inferior region of the globus pallidus (GP), and lateral to the red nucelus (RN), on the superiormedial edge of the substantia nigra (SN). The RN is a circular structure, while the GP is divided into the internal (GPi) and external (GPe) parts, and is situated directly medial to the putamen (Pu). Note, however, the level in which the RN can be seen is too far caudal to visualize the GPi on the same coronal image.
2.1 Locate the STN on the MR image
• Use the coronal view.
Figure 1. Location of the STN and surrounding structures taken from the MAI atlas. R = red nucleus, SNC & SNR = substantia nigra, STn = subthalamic nucleus, EGP
= globus pallidus external, PU = putamen. The GPi sits medial to the GPe, but cannot be seen at this level. Arrows point to location of these structures. Note, red
Locate the RN, which can be found medial to the inferior region of the PU, as can be seen in figure 2. If this structure is visible, the SN will curve around its inferior and lateral edge, separated by the ventral tegmental area (VTA). The further caudal the slice is, less of the GP • will be visible. In fact, in some scans it may be that only the Pu can be seen at this level (in
figure 2 the GPe is visible, this might not always be the case).
• The RN lies posterior to the STN within the brain, therefore once this is located, move in a rostral direction.
• Notice the STN, which will begin to appear on the superior medial edge of the SN (figure 3a). • Move further rostral, and the superior-lateral edge of the STN will start to separate further
from the SN. Concurrently, the RN will reduce in size (figure 3b).
• Eventually, the RN will no longer be visible, leaving only the SN and STN. The STN will take the shape of an ellipse (figure 4).
Figure 2.
2.2 Identify the STN on all three view-points
• Once you can see the STN on the coronal image, locate it in the other two views by selecting it with the ‘curser mode’ button. Zoom in on all three viewpoints (Figure 5, sagittal (left), coronal (middle), and axial (right)). Green cross indicates where the STN is, click the ‘toggle cross hair’ button if you would like to see the green cross.
3.0 Start segmentation
• Use the coronal view to navigate to the slice in which the STN is largest, and the STN/SN boundary is clearly visible (figure 6, red arrow indicates the boundary). This slice should be used as a starting point.
Figure 4.
• Confirm the STN/SN border in the sagittal view (figure 7).
• Note: it can be difficult to establish this border in the axial view to begin with.
• In the coronal view, start segmenting by first delineating as much of the lateral border between the STN and internal capsule (IC) as possible (figure 8a). If it is unclear whether a voxel should be included in the border or not, use the ‘curser mode’ and ‘toggle cross hair button’ to check it from all three views. Figure 8b shows this border from the sagittal view, while figure 8c shows it from the axial view. Check whether the location of the voxel is consistent with the shape of the STN. If it seems to fit in all three views, then include it. If you can see from a certain view that it does not, leave it out. This technique should be used throughout segmentation for all borders.
Figure 6.
• If possible, continue to segment the inferior lateral border between the STN and SN (figure 9).
• Note, however, that segmentation may become more difficult when moving towards this border, as the SN touches the bottom of the STN. Thus, due to partial volume effects, it may be the case that some voxels belong to both structures. It is therefore often difficult to exactly determine the border between them, as in figure 10a (red arrow indicates generally where the border should be).
• If this is the case, segment as much of the border as possible (figure 10b).
IC
IC
Figure 8a. Figure 8b. Figure 8c. Figure 9.• Subsequently, use the sagittal view to see whether it can be determined better, as in figure 11 (red arrow indicates the border). If not, leave it and continue to segment other borders which are clear in the current slice. Continuing to segment using other slices should eventually help to determine the rest of this border.
• Next, segment the posterior-medial border between the STN and zona incerta (ZI) (figure 12a). Again, if the border is not clear in the coronal view, use check it from both the sagittal (figure 12b) and axial views (figure 12a). If it is still not clear, leave it. Segmenting in other slices should help to determine the rest of the border.
Figure 10a. Figure 10b.
4.0 Continue from the start point
• Now move in a caudal direction, and segment on all slices in which the STN is visible. In each slice, segment the lateral border between the STN and IC to begin with, along with as much of the STN/SN border as possible. Follow this by delineating the medial boundary between the STN and ZI.
• Again, use all three view-points to segment, depending on which displays the border you are segmenting best. This will vary between scans.
• Section 4.1 will describe how the surrounding landmarks will change when moving caudal in the coronal view, until the STN is no longer visible. Use this as a guide to determine when segmenting should end in this direction.
4.1 Rostral to caudal
• Move caudal from the start point, and notice the RN start to appear directly medial to the STN (figure 13a).
• Move further caudal, and the RN will increase in size, while STN simultaneously decreases in size (figure 13b). Figure 12a. Figure 12b. Figure 12c. Figure 12a. Figure 12c.
• Note, at this point, it may become more difficult to determine the exact shape of the STN, and the borders may be less clear than at the start point.
• When this is the case, again use the ‘curser mode’ and ‘toggle cross hair’ buttons to check each voxel which you are unsure about from all three view-points.
• Figure 14a shows example of an unclear voxel in the coronal view, 14b shows the same voxel in the sagittal view, and 14c from the axial.
• Using all the information from each view, make an informed decision on whether to include the voxel(s). Below, the border can be drawn in the axial view (figure 15d). The most appropriate view will vary between scans and borders.
Figure 13a. Figure 13b.
Figure 14a. Figure 14b.
• Continue moving caudal and segment the STN until the RN is at its largest. Around this slice the STN will no longer visible. This will leave only the RN and SN on display (figure 15). • Stop segmenting in the caudal direction when this is the case.
4.2 Check the shape of the STN
• Use the coronal view and run through all of the slices which have been segmented so far. Check the STN has the shape of an ellipse (the shape may not be perfect), use figure 17 as a reference (from left to right moving further caudal). Note figure 16 is only 4 slices as an example, while the STN is visible in more than this.
• Now move back to the start point, and move in a rostral direction, segmenting on all the slices in which the STN is visible in this direction. Use the same order of borders as in the caudal direction.
4.3 Caudal to rostral
• As you move rostral, notice the IC encapsulate more of the lateral border of the STN (figure 17).
Figure 15.
• Move further rostral, and look out for the mamillary body (MB), a circular structure located medial and inferior to the STN (in-line with the SN) (figure 18a, red arrow points to MB). • Note, however, the MB are not always as well defined as in figure 18a. Sometimes they
resemble more a unified region of hyperintensity, rather than two individual structures (figure 18b, red circle indicates where MB are). Nevertheless, this region of hyperintensity can still be used as a landmark.
• Simultaneously, note that both the STN and SN start to reduce in size.
• Similar to when moving caudal, as you move further rostral and the STN decreases in size, the shape of the STN may be less clear. When this is the case, use the same method as shown in section 4.1, by using all three views to check accuracy.
• Note, also, when moving rostral and reaching the anterior tip of the STN that it may be unclear where to define this border, as it can seem to blur into the GP. This issue normally affects the SN more, however, it should be noted for the STN as well.
• Below, this can be seen in the sagittal (figure 19a) and axial views (figure 19b). Figure 17.
• To define this border, once again use the toggle cross hair button to check from all three views (figure 20, left sagittal, middle coronal, right axial). In the image below, from the sagittal view it is quite unclear where to draw the boundary. It can be confirmed more accurately in the coronal and axial view.
• Move further rostral and segment until the STN and SN are no longer visible, while the MB show as a hyperintense circular structure (Figure 21a). Again, it may be that the MB do not show as two independent structures, but more a region of hyperintensity. However, if this is visible and the STN and SN can no longer be seen, end segmenting in the rostral direction (figure 21b).
Figure 19a. Figure 19b.
Figure 20.
4.4 Check the shape of the STN
• Check the STN has the shape of an ellipse in all slices segmented in the rostral direction. Refer to figure 22 when doing this (from left to right moving further rostral).
5.0 Check shape in all views
• Now run through all of the slices in the sagittal and axial views points which have been used for segmentation. You may notice that certain borders are missing and need filling in (figure 23a, displays the sagittal view, 23b, displays the axial view).
• Fill in the missing borders (figure 24c, 24d). Figure 22.
Figure 23a.
Figure 23b.
6.0 End segmentation
• Fill in all segmented slices which are not yet filled with the ‘fill tool’.
• Run through all of the dimensions to ensure there are no gaps, or parts which look extremely out of place.
• Take a look at the three dimensional (3D) image. o Tools > 3D viewer
• Once 3D, the STN should resemble an almond (figure 25). If not, go back through the slices and check there are no areas which seem out of place.
• Once checked, save the mask o File > save as.
Substantia Nigra Segmentation protocol
Note this protocol will focus on the left substantia nigra (SN). The same segmentation procedure applies to the right SN. Segment the left and right SN separately, creating individual masks for each.
1.0. Start-up fsl
• Open fslview
• MATE Terminal • Type ‘fslview’.
• Choose the QSM image to segment
• File > open > home/public/HumanAtlas/subcortex/R01/
subcortex_scannumber_subjectnumer_R01_INV2_e4_P_QSM).
• Adjust the contrast so that you can see the definition between different structures; around -0.3/0.3 is a good guide. However, what is optimal may vary between scans.
• Create a mask
• File > create mask. 2.0. Refer to the MAI atlas
•
location of the SN and structures surrounding it. All structures mentioned will appear as hyperintense on QSM images.
• The SN is a curvilinear structure, divided into two components: the substantia nigra pars compacta (SNp) and substantia nigra pars recticular (SNr). Note, however, this protocol will focus on segmenting the entire SN, and not separating it into two parts. It is located inferior and medial the globus pallidus (GP), and lateral to the mamillary body (MB). The GP is divided into the internal (GPi) and external (GPe) parts, and is situated medially to the putamen (Pu). The MB is a circular structure, medial and inferior to the GP. The inferior tip if the SN is around the same level of the MB, while the superior region is slightly above. The SN sits at an oblique angle, with the superior tip further lateral to the MB than the inferior region.
2.1 Locate the SN on the MR image
• Use the coronal view.
atlas. MM = mammillary body, SNR = substantia nigra, IGP = globus pallidus internal, EGP = globus pallidus external, Pu = putamen. Arrows point to the
location of these structures. Note, red arrow points towards the SN.
Use the following information, combined with figure 1, to become familiarized with the
• Locate the MB, as can be seen in figure 2a. If this structure is visible, the GPe, Pu and most of the GPi should also be seen.
•
resemble more a unified region of hyperintensity, rather than two individual structures (figure 2b, red circle indicates where MB are). If this is the case, it is still possible to use this region of hyperintensity, combined with the GP and Pu, to locate the SN.
• The MB is located anterior to the SN within the brain, therefore move in a caudal direction. • Directly lateral to the MB, notice the SN appear. On the superior medial edge of the SN, the
subthalamic nucleus (STN) will also start to show (Figure 3a).
• Move further caudal so that the STN moves in a lateral direction, and its superior lateral border separates further from the SN. Concurrently, if the MB were displayed as two
individual circles, they will now become less defined, and seem to merge into one (figure 3b, red circle indicates where MB are).
• Stop moving caudal when the SN and STN are both clearly visible (Figure 4a shows this from the MAI atlas, STh = STN. Figure 4b shows this from the QSM image).
Figure 3a.
Figure3b.
Note, however, that the MB are not always as well defined as in figure 2a. Sometimes they
2.1 Identify the SN on all three view-points
• Once you can see the SN on the coronal image, locate it in the other two views by selecting it with the ‘curser mode’ button. Zoom in on all three viewpoints (Figure 5, sagittal (left), coronal (middle), and axial (right)). Green cross indicates where the SN is, click the ‘toggle cross hair’ button to see the green cross.
4.0 Start segmentation
• Use the coronal view to navigate to the slice in which the STN is largest, and the STN/SN boundary is clearly visible (figure 6a, red arrow indicates the boundary). This slice should be used as a starting point.
• Confirm the STN/SN border in the sagittal view (Figure 6b).
• Note: it can be difficult to establish this border in the axial view to begin with. Figure 4a.
Figure 5.
Figure 6a. Figure 6b.
• In the coronal view, start segmenting by first delineating as much of the lateral border between the STN and cerebral peduncle (CP) as possible. You may find the superior lateral border to have a jagged appearance due to the comb system. The CP forms a continuation of the internal capsule (IC) (figure 7a).
• If it is unclear where to draw the border in the coronal view, check it from the axial (figure 7b) and sagittal (figure 7c) view-points.
• If it is unclear whether to choose one particular voxel, select the ‘curser mode’ button, along with the ‘toggle cross hair’ button, and select the voxel. See an example below in figure 8 (figure 8a coronal view, 8b sagittal, 8c axial). Check whether the location of the voxel is consistent with the shape of the SN in all three views. If it seems to fit in all three, then include it. If you can see from a certain view that it does not, leave it out. This technique should be used throughout segmentation for all borders.
SN
comb
CP
IC
SN
CP
SN
CP
Figure 7a. Figure 7b. Figure 7c.IC
CP
Figure 7a.• In this example, the border can be drawn in the axial view (figure 9).
• Continue by segmenting the posterior-medial border between the SN and ventral tegmental area (VTA) (Figure 10).
• Again, if it is unclear in the coronal view, try from the sagittal (figure 10a) or axial (figure 10b). Figure 8a.
Figure 8b.
Figure 8c.
• If possible, continue to segment the superior medial border between the SN and STN (figure 11a).
• Note, however, that segmentation may become more difficult when moving towards this border, as the SN touches the bottom of the STN (Figure 11b). Thus, due to partial volume effects, it may be the case that some voxels belong to both structures. It is therefore often difficult to exactly determine the border between them
• If this is the case, segment as much of the border as possible (figure 12a).
VTA
VTA
VTA
Figure 10a. Figure 10a. Figure 10b.Figure 11a. Figure 11b.
VTA
• Subsequently, use the sagittal view to see whether it can be determined better (figure 12b, red arrow indicates where the border is). If not, continuing to segment using other slices should eventually determine the rest of this border.
4.0 Continue from the start point
• Now move in a caudal direction, and segment on all slices in which the STN is visible. In each slice, segment the lateral border between the STN and CP to begin with. Follow this by delineating the medial boundary between the STN and VTA, along with as much of the STN/SN border as possible.
• Again, use all three view-points to segment, depending on which displays the border you are segmenting best. This will vary between scans.
• Section 4.1 will describe how the surrounding landmarks will change when moving caudal in the coronal view, until the SN is no longer visible. Use this as a guide to determine when segmenting should end in this direction.
4.1 Rostral to caudal
• Move caudal from the start point, and notice the red nucleus (RN) - a circular structure located above the posterior region of the SN - start to appear directly medial to the STN (figure 13).
• Move further caudal, so that the RN increases in size, while STN simultaneously decreases in size.
• The SN will remain visible, curving around the inferior lateral aspect of the RN.
• Continue moving caudal while the RN increases in size, and the STN reduces. Notice that once the RN is at its largest, the STN is normally no longer visible at this level.
Figure 12a. Figure 12b.
• At this point, when you start to reach the posterior third of the SN, you may find that some of the voxels within the SN are of lower intensity, seeming to split the SN (figure 14a). In the axial view this may resemble a swallow tail, and in the sagittal view it may seem a whole section is missing (figure 14b).
• This is most likely a result of nigrosome 1, a small cluster of dopaminergic cells within the SN. It should be noted, and included in the mask, as it is part of the SN (Schwarz et al., 2014). Figure 14c now shows how figure 9b should be segmented, for example.
• Furthermore, around this level ensure you do not include the RN in the inferior medial tip of the SN. As you can see from Figure 15, in some slices it can be unclear which voxels are part of which structure. If this is the case, define the border as much as possible. Then check each individual voxel you are unsure about as described in section 3.0. If it is still not clear, then leave the border in this slice and continue segmenting the other borders which are clear. Segmenting in the rest of the slices should eventually determine the rest of the border. Figure 14a.
Figure 14c.
Figure 15. Figure 14b.
