Methods for cerebellar imaging analysis

The human cerebellum consists of an extended, highly-folded, and laminated cortical sheet overlying the white matter and cerebellar nuclei. This complex anatomy hinders the processing of magnetic resonance imaging data, as the relevant structures are not fully resolved. In this review, we explore the typical processing techniques employed for the anatomical and functional imaging of the cerebellum, along with their primary limitations with respect to imaging fidelity. Moreover, we discuss emerging methods applicable postmortem and in-vivo that greatly enhance fidelity in cerebellar imaging through higher-resolution data and individual-level processing


Challenges in human cerebellar imaging
Similar to the cerebrum, processing of MRI data requires image enhancement, the delineation of the relevant areas of interest, such as the gray and white matter, and potentially the application of additional transformations, such as group normalization or surface projection.These steps, however, are particularly challenging in the cerebellum, since the anatomy is typically not fully resolved, due to the small size of the relevant structures, that is the cerebellar nuclei and thin cortex (Figure 1a-b [4]).The resulting partial volume effects, as well as the tight foliation of the cerebellar cortex, hinder the separation of gray matter banks, particularly with segmentation pipelines that are optimized in the cerebrum.Cerebellar segmentation is further hindered by adjacent structures, such as the dura and sinuses, that, depending on the MRI sequence, may have a similar signal intensity to the cortex, while the fine cerebellar vasculature and cortical folding are characterized by similarly-high spatial frequencies which render edge-detection challenging.Furthermore, registration-based approaches, for example, to achieve segmentation or group homogenization are severely limited by the individual-dependent complex cerebellar foliation, which reduces registration accuracy beyond the main cerebellar lobules.The complex anatomy also hinders the creation of well-defined segmentations and surfaces (i.e.without holes/handles), while the inflation of these highlycurved surfaces results in large distortions [4].All these are exacerbated by the relatively low signal-to-noise (SNR) and high signal inhomogeneity in the cerebellum with typical coil geometries, particularly in ultra-high-field MRI [5], as well as pulsatility artifacts close to the 4th ventricle.Finally, even if adequately high-resolution datasets are acquired, most neuroimaging packages struggle with the increased computational requirements since they are optimized for relatively low-resolution datasets [6].The above challenges greatly hinder cerebellar research resulting in the cerebellum being relatively understudied in-vivo.

Image enhancement
Image enhancement techniques can help ameliorate some of these issues in structural images: intensity-bias correction is a common first step to homogenize the signal distribution [7].

Denoising approaches can ]] ]] ]]]]]]
improve SNR of high-resolution data, for example by leveraging the redundancy of multi-dimensional MRI data.Such approaches may be particularly useful for the lower-SNR later echoes of multi-echo data used to visualize the cerebellar nuclei [8].Super-resolution acquisition/reconstruction approaches can also help improve the SNR/resolution tradeoff in the cerebellum, as shown recently in diffusion data [9].

Tissue classification
Tissue classification in the cerebellum is commonly performed through neuroimaging packages like the cerebellum-specific SUIT [12], CERES2 [13], and ACAPULCO [14], or the cerebrum-focused Freesurfer [15] (for a performance evaluation, see [10], Figure 1c).These packages rely on atlases or combined atlas and signal-intensity automated methods, thus segmenting the cerebellar lobules in a reproducible manner.Similar approaches exist for the cerebellar nuclei [3,12,16,17].These methods therefore greatly facilitate the analysis of lower resolution data, such as from functional MRI (fMRI) or non-specialized cerebellar imaging approaches [12], however, they implicitly segment at a spatial scale that smooths over the fine anatomical detail, such as the individual cerebellar folia, or the dentate cortex.This is a pragmatic choice since the current in-vivo state-of-the-art MRI scans do not fully resolve the cerebellar anatomy and the computational demands increase quickly with finer sampling (Figure 1d-e).
These streamlined methods have therefore found widespread usage in human neuroscience, but produce cortical measures (e.g.cortical thickness) several factors off of the biological ground-truth [1,4].Recently-released packages, such as the Nighres platform [18] and the submillimeter CHROMA atlas (Figure 1f, Figure 2a-b, [19,20]), may better approximate the cerebellar cortical surface with increasing image resolution.Cerebellar nuclei parcellation may also benefit from more elaborate shape-based methods at higher resolution, for example [21,22].

Individual-level segmentation
The limited resolving power compared to the fine cerebellar anatomy, stemming from acquisition and analysis limitations, has hindered even post-mortem research, with the large-scale cerebellar cortical features being less well-characterized compared to the cerebrum.Only recently the accurate reconstruction of the complete cerebellar cortical surface was achieved from post-mortem MRI [4] or 3D-reconstructed histology data [11].Both of these studies used high-resolution data combined with signal-intensity-based methods and manual corrections to segment the cerebellar surface to the level of individual folia.They showed that the cerebellar cortical sheet has a similar surface area to the cerebral, though high variability remains in the thickness estimates of the granular and molecular layers across the cerebellum [11,23].We recently showed that dedicated cerebellar (a) (bottom; adapted with permission from [4]).(c) Adapted with permission from [10].(d) adapted with permission from [11].
imaging at 7T combined with motion correction, denoising, and semi-automated intensity-based segmentation algorithms can closely approximate the cerebellar surface in vivo, though still lagging compared to these post-mortem datasets (Figure 2c-d, [1]).Acquisition and analysis improvements to streamline cerebellar segmentation in-vivo while retaining anatomical detail remains an important target for cerebellar imaging.

Functional MRI in the cerebellum
Functional MRI of the cerebellum has convincingly shown its importance across the cognitive domain [24][25][26][27][28][29].Compared to fMRI in the cerebral cortex, cerebellar fMRI further suffers from increased susceptibilityinhomogeneities and physiological noise.With typical fMRI acquisition techniques and particularly at higher fields, this may require distortion correction and physiological noise removal techniques to be applied during preprocessing [30].Depending on the B 0 inhomogeneity change throughout the acquisition, non-linear motion correction techniques may also be required.The projection of fMRI results to reference cerebellar atlases [12,31], and associated surface-based representations [12,32,33] has greatly facilitated functional cerebellar research in humans [26,34,35].Recently, the combination of accurate structural reconstruction of the cerebellum and high-resolution fMRI has allowed the detailed examination of the functional activation of the cerebellar cortex [20], revealing a similar somatotopic organization to the cerebral cortex.Note that the functional organization of the cerebellum is not limited to specific lobules [26]: electrophysiology animal studies indicate instead a finer patch and stripe-like organization [36].Currently, the blood oxygen level dependent (BOLD) fMRI effective resolution limits have not allowed studying this finer functional organization in the cerebellum, but the recent advent of high-specificity fMRI techniques may allow this [37].This may require adapting the cerebellar fMRI processing pipelines to facilitate examining more spatially-specific activation, similar, for example to layerspecific analysis in the cerebral cortex [32,[38][39][40].

Conclusions
Overall, many of the image processing tools needed for detailed analysis of cerebellar structure and function are available, from denoising and super-resolution all the way to cortical mapping.The main challenge still resides in the cerebellar incredibly compact organization, which requires stretching the capabilities of MR imaging to the limit.However, these are rapidly changing, combining optimized acquisition techniques, multi-dimensional contrasts, and machine learning reconstruction to push the boundary of in-vivo MRI [1,[41][42][43] while the latest efforts of post-mortem MRI and whole brain microscopy provide invaluable data sets to start scaling up analysis pipelines to account for individual folia and nuclei in the near future [4,11,23,44,45].(c) (adapted with permission from [1]).(adapted with permission from [4]).

Figure 1 Current
Figure 1

Figure 2 Current
Figure 2