A standardized pipeline is presented for examining cerebellum grey matter morphometry. The pipeline combines high-resolution, state-of-the-art approaches for optimized and automated cerebellum parcellation and voxel-based registration of the cerebellum for volumetric quantification.
Multiple lines of research provide compelling evidence for a role of the cerebellum in a wide array of cognitive and affective functions, going far beyond its historical association with motor control. Structural and functional neuroimaging studies have further refined understanding of the functional neuroanatomy of the cerebellum beyond its anatomical divisions, highlighting the need for the examination of individual cerebellar subunits in healthy variability and neurological diseases. This paper presents a standardized pipeline for examining cerebellum grey matter morphometry that combines high-resolution, state-of-the-art approaches for optimized and automated cerebellum parcellation (Automatic Cerebellum Anatomical Parcellation using U-Net Locally Constrained Optimization; ACAPULCO) and voxel-based registration of the cerebellum (Spatially Unbiased Infra-tentorial Template; SUIT) for volumetric quantification.
The pipeline has broad applicability to a range of neurological diseases and is fully automated, with manual intervention only required for quality control of the outputs. The pipeline is freely available, with substantial accompanying documentation, and can be run on Mac, Windows, and Linux operating systems. The pipeline is applied in a cohort of individuals with Friedreich ataxia (FRDA), and representative results, as well as recommendations on group-level inferential statistical analyses, are provided. This pipeline could facilitate reliability and reproducibility across the field, ultimately providing a powerful methodological approach for characterizing and tracking cerebellar structural changes in neurological diseases.
The cerebellum is a part of the brain historically associated with motor control1,2,3 and is thought to be integrally involved in only a small set of rare diseases, such as inherited ataxias4. However, converging lines of research from anatomical tracing studies in nonhuman primates, as well as human lesion and neuroimaging studies, provide compelling evidence for a role of the cerebellum in a wide array of cognitive5,6,7, affective8,9,10,11, and other nonmotor functions7,12 (see6 for review). Furthermore, abnormalities of the cerebellum are increasingly implicated in a broad range of neurological and psychiatric disorders, including Parkinson's disease13, Alzheimer's disease14,15, epilepsy16,17, schizophrenia18, and autism spectrum disorder19. Therefore, it has become essential to incorporate the cerebellum into functional and structural models of human brain diseases and normative behavioral variability.
Anatomically, the cerebellum can be divided along its superior to inferior axis into three lobes: anterior, posterior, and flocculonodular. The lobes are further subdivided into 10 lobules denoted by roman numerals I-X20,21 (Figure 1). The cerebellum can also be grouped into midline (vermis) and lateral (hemisphere) zones, which respectively receive inputs from the spinal cord and cerebral cortex. The anterior lobe, comprising lobules I-V, has traditionally been associated with motor processes and has reciprocal connections with cerebral motor cortices22. The posterior lobe, comprising lobules VI-IX, is primarily associated with nonmotor processes11 and has reciprocal connections with the prefrontal cortex, posterior parietal, and superior temporal cerebral cortices8,23. Lastly, the flocculonodular lobe, comprising lobule X, has reciprocal connections with vestibular nuclei that govern eye movements and body equilibrium during stance and gait21.
A growing body of recent work using functional neuroimaging has further refined understanding of the functional neuroanatomy of the cerebellum beyond its anatomical divisions. For example, resting-state functional magnetic resonance imaging (fMRI) techniques have been used to map the pattern of functional interactions between the cerebellum and cerebrum24. Additionally, using a task-based parcellation approach, King and colleagues7 demonstrated that the cerebellum shows a rich and complex pattern of functional specialization across its breadth, evidenced by distinct functional boundaries associated with a variety of motor, affective, social, and cognitive tasks. Collectively, these studies highlight the importance of examining individual cerebellar subunits to develop complete biological characterizations of cerebellum involvement in both healthy variability and neurological diseases characterized by alterations in cerebellar structure and/or function.
The present work focuses on methods for quantifying local changes in cerebellar volume using structural MRI in humans. In general, there are two fundamental approaches to the quantification of regional brain volume using MRI data: feature-based segmentation and voxel-based registration. Feature-based segmentation approaches use anatomical landmarks and standardized atlases to automatically identify boundaries between subregions. Mainstream software packages for segmentation include FreeSurfer25, BrainSuite26, and FSL-FIRST27. However, these packages provide only coarse parcellations of the cerebellum (e.g., labeling the whole grey matter and whole white matter in each hemisphere), thus overlooking the individual cerebellar lobules. These approaches are also prone to mis-segmentation, particularly overinclusion of the surrounding vasculature.
New machine-learning and multi-atlas labeling algorithms have been developed, which provide more accurate and finer-grained parcellation of the cerebellum, including Automatic Classification of Cerebellar Lobules Algorithm using Implicit Multi-boundary evolution (ACCLAIM28,29), Cerebellar Analysis Toolkit (CATK30), Multiple Automatically Generated Templates (MAGeT31), Rapid automatic segmentation of the human cerebellum and its lobules (RASCAL32), graph-cut segmentation33, and CEREbellum Segmentation (CERES34). In a recent paper comparing state-of-the-art fully automated cerebellum parcellation approaches, CERES2 was found to outperform other approaches relative to gold-standard manual segmentation of the cerebellar lobules35. More recently, Han and colleagues36 developed a deep-learning algorithm called ACAPULCO (Automatic Cerebellum Anatomical Parcellation using U-Net with locally constrained optimization), which performs on par with CERES2, has broad applicability to both healthy and atrophied cerebellums, is available in open-source Docker and Singularity container format for 'off-the-shelf' implementation, and is more time-efficient than other approaches. ACAPULCO automatically parcellates the cerebellum into 28 anatomical regions.
In contrast to feature-based segmentation, voxel-based registration approaches operate by precisely mapping an MRI to a template image. To achieve this mapping, the voxels in the original image must be distorted in size and shape. The magnitude of this distortion effectively provides a measure of volume at each voxel relative to the gold-standard template. This form of volumetric assessment is known as 'voxel-based morphometry'37. Whole-brain voxel-based registration approaches, such as FSL-FLIRT38/FNIRT39, SPM unified segmentation40, and CAT1241, are commonly used for voxel-based morphometry. However, these approaches do not account well for the cerebellum, resulting in poor reliability and validity in infratentorial regions (cerebellum, brainstem42). To account for these limitations, the SUIT (Spatially Unbiased Infra-tentorial Template) algorithm was developed to optimize cerebellum registration and improve the accuracy of voxel-based morphometry42,43.
Feature-based segmentation and voxel-based registration approaches for the estimation of regional cerebellar volume have fundamental strengths and weaknesses. Segmentation approaches are substantially more accurate for quantifying the volume of anatomically defined areas (e.g., lobules35). However, boundaries between distinct functional modules of the cerebellum do not map onto its anatomical folia and fissures (equivalent to gyri and sulci of the cerebrum7). As registration-based approaches are not constrained by anatomical landmarks, finer-grained spatial inference and high dimensional structure-function mapping of the cerebellum is possible44. Taken together, segmentation and registration approaches are complementary to one another and can be used to answer different research questions.
Here, a new standardized pipeline is presented, which integrates these existing, validated approaches to provide optimized and automated parcellation (ACAPULCO) and voxel-based registration of the cerebellum (SUIT) for volumetric quantification (Figure 2). The pipeline builds upon the established approaches to include quality control protocols, using qualitative visualization and quantitative outlier detection, and a rapid method for obtaining an estimation of intracranial volume (ICV) using Freesurfer. The pipeline is fully automated, with manual intervention only required for checking the quality control outputs, and can be run on Mac, Windows, and Linux operating systems. The pipeline is freely available with no restrictions of its use for noncommercial purposes and can be accessed from the ENIGMA Consortium Imaging Protocols webpage (under "ENIGMA Cerebellum Volumetrics Pipeline"), following the completion of a brief registration form45.
All required software is listed in the Table of Materials, and detailed tutorials, including a live demonstration, are available upon download of the pipeline, in addition to the protocol described below. Finally, representative results are provided, from the implementation of the pipeline in a cohort of people with Friedreich ataxia (FRDA) and age-and sex-matched healthy controls, alongside recommendations for group-level statistical inferential analyses.
NOTE: The data used in this study were part of a project approved by the Monash University Human Research Ethics Committee (project 7810). Participants provided written informed consent. While the pipeline can be run on Mac, Windows, or Linux operating systems, ACAPULCO, SUIT, and the QC pipelines have explicitly been tested on Linux (Ubuntu) and Mac (Catalina, Big Sur v11.0.1) operating systems.
1. Module 1: ACAPULCO (anatomical parcellation)
2. Module 2: SUIT cerebellum-optimized voxel-based morphometry
3. MODULE 3 (optional): Intracranial Volume (ICV) estimation using FreeSurfer
NOTE: This module will use the FreeSurfer pipeline to calculate ICV. It does not need to be re-run if there are existing Freesurfer outputs for the cohort (any version).
Cerebellum parcellation (ACAPULCO)
Quality control of cerebellum parcellated masks:
The following examples demonstrate the ACAPULCO parcellated outputs and guide decision-making about a) the quality of the parcellated mask at the individual level and b) subsequent inclusion or exclusion of a particular lobule(s) from the statistical analyses. Ultimately, the decision to include or exclude a subject is subjective; examples of 'good parcellations,' 'subtle parcellation errors,' and 'global failures' from a variety of healthy and clinical groups are provided here.
Examples of 'good parcellations' are presented in Figure 4, including in healthy and heavily atrophied cerebella. In Figure 5, subtle over- and under-inclusions of individual cerebellum lobules are depicted. These are the most common type of parcellation error and may not be detected as statistical outliers in the quantitative QC protocol. These types of errors generally require the exclusion of the individual lobules that are affected, while the remainder of the parcellated cerebellum is unaffected and can be retained. In contrast, 'Global failures,' as depicted in Figure 6, require complete exclusion of the subject.
Statistical outlier detection:
To illustrate the pipeline, ACAPULCO was run on a sample of 31 people with FRDA (mean age= 36.5 years; SD= 13.0 years, 14 females) and 37 age- and sex-matched healthy controls (HC) (mean age= 37.1 years; SD=12.8 years, 17 females) as previously described55. Across the whole sample, 18 lobules were detected as statistical outliers (<1% of the total sample). After performing detailed QC on the images, 17 outlier lobules were removed from group-level analyses by removing the individual lobule volume for the respective subject(s) from the group-level cerebellar volumes file (i.e., the 'Cerebel_vols.csv' file). The remaining outlier was deemed not a segmentation error but rather due to variability in the individual's cerebellum anatomy and was therefore retained in the analysis. There were also two global parcellation failures (1 FRDA patient). The base rate of exclusion across all cerebellum lobules (i.e., global parcellation failures) was 1.5%. Table 1 shows the exclusion rates for each of the 28 anatomical ROIs. Left lobule IX and Right lobule Crus I had the highest rates of exclusion.
Group-level statistical analysis:
A total of 66 subjects (30 FRDA patients) were included in the group-level analysis. Two-tailed Mann-Whitney independent samples tests were conducted to test for significant differences in cerebellum lobule volumes between FRDA and HC. Results showed significantly reduced white matter in the corpus medullare in FRDA vs. HC (p < 0.05, Bonferroni corrected for 28 comparisons). There were no other significant between-group differences. See Supplemental Table S1 for the volumes of all 28 cerebellar subunits in the sample.
Cerebellum voxel-based morphometry analyses (SUIT)
Quality control:
Examples of well-aligned images and examples of exclusions for both healthy controls and clinical groups, including FRDA and Spinocerebellar ataxia pateints, are shown in Figure 7. SUIT analyses were run on a total of 64 subjects (28 FRDA; 36 HC) from the sample described above, following the exclusion of an additional two subjects due to incomplete whole cerebellar coverage in the cerebellum mask.
After testing the spatial covariance of all normalized images in the sample relative to each other, two scans were detected as statistical outliers based on their mean spatial covariance with the rest of the sample (Figure 8). However, visual inspection of both the original and normalized images indicated that although these individuals had some unique neuroanatomy, there were no significant artifacts in either image and the processing steps completed normally. As such, both subjects were retained in the analyses.
Group-level statistical analysis:
Images were smoothed using a Gaussian kernel of 3 mm full width at half maximum (FWHM). Nonparametric permutation tests were carried out in SnPM to test for significant between-group differences in cerebellum grey matter volumes. To this end, 5,000 permutations were run, with a cluster-forming threshold of p < 0.001. Images were explicitly masked with the SUIT grey matter atlas to limit inference to grey matter regions. To correct for head size, intracranial volume was entered as a covariate in the model. Final inference of group results was carried out at p < 0.05, family-wise error (FWE) cluster-level corrected.
Compared to HC, FRDA showed significantly reduced grey matter volume in bilateral anterior lobules I-V (left: x= -10, y= -46, z= -26; T= 5.61; Ke= 754; right: x= 10, y= -38, z= -21; T= 6.83; Ke= 569); and in medial posterior lobe regions, including Vermis VI, extending bilaterally into Lobule VI (x= 3, y= -65, z= -20; T= 7.25), and Vermis IX extending bilaterally into Lobule IX (x= 3, y= -65, z= -20; T= 6.46; Ke= 3974; Figure 9).
Figure 1: The human cerebellum. (A) A flattened representation of the cerebellum and its major fissures, lobes, and lobules. Red = anterior lobe (lobules I-V); cream= posterior lobe (lobules IV-IX); purple= flocculonodular lobe. The cerebellum can be divided into midline "vermis" and lateral "hemisphere" zones. All lobules are identified in the vermis and hemispheres. In lobule VII, VIIAf in the vermis expands in the hemispheres to become Crus I, lobule VIIAt at the vermis becomes Crus II in the hemispheres, and lobule VIIB retains its label both in the vermis and hemispheres. (B) Top: Cerebellum flat map showing the anatomical subunits of the cerebellum in different colors. Bottom: a posterior view of the cerebellum. This figure was adapted from 56 and 57. Please click here to view a larger version of this figure.
Figure 2: Schematic illustration of the pipeline. A high-resolution T1 MPRAGE image is required. There are three modules: ACAPULCO, SUIT, and ICV. The pipeline is fully automated (except for manual intervention required for QC of the outputs), is available in Docker and Singularity container format, and takes approximately 20 min to run from start to finish, per subject. Please click here to view a larger version of this figure.
Figure 3: ACAPULCO parcellation demonstrating each of the 28 anatomical subunits. This figure was adapted from36. Abbreviations: CM = corpus medullare; Ver = vermis; R/L = right/left. Please click here to view a larger version of this figure.
Figure 4: Examples of 'good' cerebellum parcellations. Sagittal (x= 0) and coronal (y= -57) slices are shown. (A, B) Sagittal and coronal slices showing parcellated masks from two healthy cerebellar. The algorithm has localized the individual subunit boundaries correctly, and it has not overincluded the transverse sinus into the labeling and quantification of Crus I. (C) A heavily atrophied cerebellum of a SCA2 patient. Here, atrophy is apparent along the entire extent of the cerebellum, the sulci are broad, and there is a lot of missing tissue. There is a slight overinclusion of the vasculature in the posterior lobes that is more pronounced on the right side (yellow arrow). Aside from this overinclusion, ACAPULCO has worked well. Abbreviations: CM = corpus medullare; Ver = vermis; L/R = left/right. Please click here to view a larger version of this figure.
Figure 5: Examples of cerebellar 'mis parcellations.' (A) Sagittal slice (x= 0) and coronal slice (y = -57) showing a parcellation error in an FRDA patient. The algorithm has not worked well along the midline, and underinclusions of Crus I and II (red arrow) are evident along the posterior extent. These lobules would be excluded from subsequent group-level analyses. (B) Sagittal slice (x= 8), coronal slice (y= -47) showing a parcellation error in a healthy cerebellar. The algorithm has completely missed left lobule VIIIb (red arrows). This lobule would be excluded from subsequent group-level analyses. (C) Sagittal slice (x= -24) and coronal slice (y= -47) showing a parcellation error in a healthy cerebellar. Some cerebellar atrophy is present, and there is an underinclusion of Crus I (red arrow). Abbreviations: FRDA = Friedreich ataxia; CM = corpus medullare; Ver = vermis; L/R = left/right. Please click here to view a larger version of this figure.
Figure 6: Examples of 'global failures' of cerebellar parcellation. (A) Sagittal slice (x= 0) and coronal slice (y= -57) showing a parcellation failure. Here, the cerebellum is only partly segmented, and parts of the occipital lobe have incorrectly been labeled as the cerebellum. These types of failures are likely due to a problem with the header of the original image, which will affect the ACAPULCO affine transformation of the image into world coordinate and subsequent localization of the cerebellum. (B) Sagittal slice (x= 0) and coronal slice (x= -57) showing a parcellation failure in a person with FRDA. Here, the CM has been completely mis-segmented. The algorithm has labeled the CM at the back of the head (red oval), outside the brain. The boundaries of the white matter have not been captured and are mislabeled as grey matter, particularly affecting left lobules VIIIb and IX. Left lobule X has also been missed (red arrow on coronal slice). These examples warrant immediate exclusion from group-level analyses for ACAPULCO and SUIT analyses. Abbreviations: CM = corpus medullare; Ver = vermis; L/R = left/right. Please click here to view a larger version of this figure.
Figure 7: Warped and modulated voxel-based morphometry maps. (A, B) Well-aligned cerebellar grey matter from two HCs. (C) A cerebellum from an HC that was detected as a statistical outlier but was retained in the analysis. (D) An atrophied cerebellum from a person with FRDA. The cerebellum has been warped correctly to the template; hence, this would not warrant exclusion. (E) An exclusion. There is a gradient in contrast from the top to the bottom of the image that reflects an error in the processing. (F) A hyperintense plane artifact in the lower right of the image of unknown origin necessitates exclusion. (G) An example of a heavily atrophied cerebellar of an SCA2 patient. The cerebellum has been warped correctly to the template; however, there is a lot of missing tissue, resulting in low contrast. This would not be an exclusion. (H) Example of poor masking necessitating exclusion. Abbreviations: VBM = voxel-based morphometry; HC = healthy control; FRDA = Friedreich ataxia; SCA2 = Spinocerebellar ataxia 2. Please click here to view a larger version of this figure.
Figure 8: Spatial covariance of SUIT voxel-based morphometry maps. Box plot illustrating the spatial covariance of the voxel-based morphometry maps for a cohort of 64 (28 FRDA) subjects. Spatial covariance is a measure of how well aligned each image is, relative to every other image in the sample. The data are tightly clustered together with an average spatial covariance correlation of ~0.95. Here, two outliers (1 FRDA, 1 HC) were detected, as >2 SDs below the mean. Abbreviations: FRDA = Friedreich ataxia; HC = healthy control; SD = standard deviation; cont = control. Please click here to view a larger version of this figure.
Figure 9. Between-group results of a voxel-based morphometry analysis of cerebellar grey matter morphometry. (A) Sagittal, (B) coronal, and (C) salt map representations of voxel-level statistical maps in individuals with FRDA vs. controls, controlling for ICV. Only voxels that survive p < 0.05 FWE cluster-level corrected, are shown. Color bar indicates T statistic. Abbreviations: FRDA = Friedreich ataxia; ICV = intracranial volume; FWE = family-wise error. Please click here to view a larger version of this figure.
Lobule | % excluded |
CM | 1.5 |
Left Crus I | 2.9 |
Left Crus II | 2.9 |
Left I III | 2.9 |
Left IV | 1.5 |
Left IX | 4.51 |
Left V | 2.9 |
Left VI | 1.5 |
Left VIIB | 1.5 |
Left VIIIA | 1.5 |
Left VIIIB | 2.9 |
Left X | 2.9 |
Right Crus I | 6.02 |
Right Crus II | 2.9 |
Right I III | 2.9 |
Right IV | 1.5 |
Right IX | 2.9 |
Right V | 1.5 |
Right VI | 1.5 |
Right VIIB | 1.5 |
Right VIIIA | 2.9 |
Right VIIIB | 1.5 |
Right X | 1.5 |
Vermis IX | 1.5 |
Vermis VI | 2.9 |
Vermis VII | 1.5 |
Vermis VIII | 1.5 |
Vermis X | 1.5 |
Table 1: Cerebellar anatomical lobules derived from ACAPULCO and rates of exclusion (%) across a sample of 31 people with FRDA and 37 HCs. Abbreviations: FRDA = Friedreich ataxia; HC = healthy controls.
Supplemental Table S1: Volumes (mm3) of 28 cerebellar anatomical lobules in Friedreich Ataxia and healthy control individuals. Please click here to download this Table.
The cerebellum is critical to a wide range of human motor3, cognitive58, affective10, and language7,59 functions and is implicated in many neurological and psychiatric diseases. The availability of a standardized and easily implementable approach for the quantification of regional cerebellar volumes will contribute to increasingly detailed 'whole-brain' structure-function mapping, complete disease modeling, and improved opportunities to define and track cerebellar contributions to brain diseases. This standardized pipeline described here combines state-of-the-art approaches for automatic cerebellum parcellation and finer-grained spatial profiling of cerebellar grey matter morphometry in both health and disease.
The results of the cross-sectional cerebellum parcellation analysis using ACAPULCO presented here showed that people with FRDA (vs. HC) had significantly reduced white matter volumes. These findings support previous studies of FRDA, which consistently show early, robust, and progressive white matter volume loss, particularly in the dentate nuclei, in FRDA. Furthermore, the pattern and extent of progressive neurodegeneration in the superior and inferior cerebellar peduncles and the dentate nuclei have been shown to differ as a factor of onset age of FRDA44. The results from the SUIT analysis revealed additional findings. Specifically, there was significant voxel-level volume loss in FRDA (vs. HC) in anterior lobe regions corresponding to bilateral lobules I-IV and Right V, extending into lobule VI. In addition, the SUIT analysis revealed significant volume loss in FRDA (vs. HC) in the medial posterior lobe regions, including lobule IX, X, and Vermis. This pattern of between-group differences is comparable to previously published work in the same cohort of FRDA patients, using a whole-brain VBM approach55.
Defining cerebellar abnormalities in neurological and psychiatric diseases is a high-priority research area with translational impact. Instrumental to tracking and treating neurological diseases-particularly those where the cerebellum is a primary site of neurodegeneration-is the development of complete biological characterizations of cerebellum involvement. The pipeline presented herein allows for relationships between individual cerebellar lobule grey matter morphometry and clinical measures that are used as the "gold standard" for clinical endpoints of disease to be explored. Such research can have a significant translational impact. For example, in the rare cerebellar diseases space, the identification of a particular profile of cerebellar grey matter atrophy in a subgroup of patients that maps onto or predicts clinical symptoms would have implications for guiding clinical practice. The inclusion of the SUIT module further allows for interesting research questions to be addressed such as structure-function mapping of the cerebellum or analysis of functional gradients of the cerebellum60.
General recommendations for group-level statistical analyses
ACAPULCO: Volumes of each cerebellar lobule (in mm3) for each subject are recorded in Cerebel_vols.csv. During statistical inference of group-level effects, intracranial volume (ICV; also recorded in Cerebel_vols.csv) should be controlled for to account for variability in head size. Alpha-significance thresholds should be corrected to account for inference across multiple lobules.
SUIT: Grey matter cerebellar VBM can be performed on the wd<subject>seg1.nii images using standard MRI processing software such as SPM or FSL. See the CAT12 manual for an excellent introduction to VBM using SPM1254. ICV should be controlled for to account for variability in head size.
For VBM in the cerebellum, it is generally recommended to use a Gaussian spatial smoothing kernel of no more than 3 mm full width at half maximum (FWHM). Appropriate statistical correction must be applied to account for multiple comparisons across voxels. In general, it is recommended to use nonparametric approaches (e.g., SnPM or FSL-Randomise).
The most critical step for successful parcellation of the cerebellum using ACAPULCO is general quality-checking of the T1 images prior a and post processing. It is highly recommended that the user check for bad contrast images (e.g., an inconsistent gradient across the image) and severe tilting of the head and motion artifacts, all of which can affect the performance of the algorithm. In addition, while the ACAPULCO algorithm has been trained on atrophied cerebellar, it has not been trained on lesion data. It is anticipated that lesions in the cerebral cortex would not be expected to impact the performance of the algorithm and subsequent accuracy of the parcellation; however, large infarcts in the cerebellum would likely yield parcellation errors. Quality-checking the cerebellum mask post processing is essential. Minor parcellation errors (e.g., minor under- and overinclusions of cerebellum lobules) are sometimes not detected as statistical outliers; conversely, instances of incorrect nonoutliers can occur where data are within the normal range, despite an obvious parcellation error. If a subject is identified as an outlier, it is essential to perform follow-up, detailed quality-checking of the cerebellum mask slice-by-slice to guide decision-making about whether to include or exclude the lobule(s) for that subject. Another critical step when running the SUIT pipeline (Module 2) is that it requires the ACAPULCO module to have already been run. Specifically, SUIT requires the cerebellum mask produced in ACAPULCO to run the cerebellum isolation and segmentation. It is important that the cerebellum mask is quality-checked to ensure full cerebellar coverage.
There are some limitations to the protocol. First, while ACAPULCO achieves state-of-the-art accuracy for cerebellar grey matter parcellation, it is not optimized for white matter parcellation; the corpus medullare covers the main body of the white matter but does not provide a measure of all the white matter. Second, the convolutional neural networks used to localize and segment the cerebellum in ACAPULCO do not generalize well to images with different contrasts or images that were not used in training. For example, because only 3T images were used in training, the parcellation quality using images acquired on a 1.5 T scanner is typically not as good; furthermore, there are no statistics with respect to ground truth that has been performed on these images. Finally, the pipeline controls for the confounding effects of head size on cerebellum volume estimates by providing an estimate of ICV that can be included as a regressor of no interest in group-level statistical analyses. However, an ideal approach would be to calculate ICV-corrected cerebellum volumes at the individual level prior to running QC, such that detected outliers reflect a true parcellation error and not natural variability in the subjects' neuroanatomy (e.g., having a large head).
In conclusion, we present a standardized pipeline for examining cerebellum grey mater morphometry, which has broad applicability to a range of neurological diseases. The pipeline is set up to allow for large, multisite studies and 'mega analyses' and is publicly available for use by research groups to facilitate reliability and reproducibility across the field. Ultimately, this pipeline provides a powerful methodological approach for further characterizing and tracking cerebellar structural changes with disease progression in neurological diseases. A longitudinal pipeline is currently being developed.
The authors have nothing to disclose.
The work presented in this manuscript was funded by an Australian National Health and Medical Research Council (NHMRC) Ideas Grant: APP1184403.
ACAPULCO pipeline files | 0.2.1 | http://enigma.ini.usc.edu/protocols/imaging-protocols/ | Please make sure to use acapulco version 0.2.1 |
Docker for Mac | https://docs.docker.com/desktop/mac/install/ | macOS must be version 10.14 or newer Docker requires sudo priviledges Docker imposes a memory (RAM) constraint on Mac OS. To increase the RAM, open Docker Desktop, go to Preferences and click on resources. Increase the Memory to the maximum |
|
Docker for Windows | https://docs.docker.com/docker-for-windows/install/ | ||
ENIGMA SUIT scripts | http://enigma.ini.usc.edu/protocols/imaging-protocols/ | ||
FreeSurfer | 7 | https://surfer.nmr.mgh.harvard.edu/fswiki/DownloadAndInstall | Following variables need to be set everytime you work with Freesurfer: export FREESURFER_HOME= ![]() ![]() source $FREESURFER_HOME/SetUpFreeSurfer.sh |
export SUBJECTS_DIR=![]() ![]() |
|||
FSL (for FSLeyes). Optional | 6 | https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FslInstallation | |
ICV pipeline files | http://enigma.ini.usc.edu/protocols/imaging-protocols/ | ICV pipeline can be run in two ways: 1) with docker/singularity. You will not require additionl software; 2) without docker/singularity- this involves running the ICV script (calculate_icv.py) manually. You will require the following additional software: Python version ![]() Python module pandas Python module fire Python module tabulate Python module Colorama |
|
https://github.com/Characterisation-Virtual-Laboratory/calculate_icv | |||
MATLAB* | 2019 or newer | https://au.mathworks.com/ | An academic license is required |
Singularity | 3.7 or newer | https://www.sylabs.io/docs/ | Prefered for high performance computing (HPC) clusters |
SPM | 12 | http://www.fil.ion.ucl.ac.uk/spm/software/spm12/ | Make sure spm12 and all subfolders are in your MATLAB path |
SUIT Toolbox | 3.4 | http://www.diedrichsenlab.org/imaging/suit_download.htm | Make sure you place SUIT toolbox in spm12/toolbox directory |
Troubleshooting manual and segmentation output examples | http://enigma.ini.usc.edu/protocols/imaging-protocols/ | ||
Tutorial manual and video | http://enigma.ini.usc.edu/protocols/imaging-protocols/ | Manual and accompanying live demonstration provide detailed step-by-step instructions on how to run the pipeline from start to finish. | |
*Not freely available; an academic license is required |