Test-retest reliability of cerebellar segmentation methods

2004 

Abstract Submission 

Ina Drabløs¹, Stener Nerland¹, Wibeke Nordhøy², Robin Bugge³, Ole Andreassen⁴, Ingrid Agartz¹, Kjetil Jørgensen⁵, Dimitrios Andreou¹

¹Department of Psychiatric Research, Diakonhjemmet Hospital, Oslo, Norway, ²Department of Physics and Computational Radiology, Division of Radiology and Nuclear Medicine, Oslo, Norway, ³Department of Diagnostic Physics, Division of Radiology and Nuclear Medicine, Oslo, Norway, ⁴NORMENT, Division of Mental Health and Addiction, Oslo University Hospital, Oslo, Norway, ⁵NORMENT, Institute of Clinical Medicine, University of Oslo, Oslo, Norway

Ina Drabløs  
Department of Psychiatric Research, Diakonhjemmet Hospital
Oslo, Norway

Stener Nerland  
Department of Psychiatric Research, Diakonhjemmet Hospital
Oslo, Norway

Wibeke Nordhøy  
Department of Physics and Computational Radiology, Division of Radiology and Nuclear Medicine
Oslo, Norway

Robin Bugge  
Department of Diagnostic Physics, Division of Radiology and Nuclear Medicine
Oslo, Norway

Ole Andreassen  
NORMENT, Division of Mental Health and Addiction, Oslo University Hospital
Oslo, Norway

Ingrid Agartz  
Department of Psychiatric Research, Diakonhjemmet Hospital
Oslo, Norway

Kjetil Jørgensen  
NORMENT, Institute of Clinical Medicine, University of Oslo
Oslo, Norway

Dimitrios Andreou  
Department of Psychiatric Research, Diakonhjemmet Hospital
Oslo, Norway

The cerebellum serves a crucial role in motor control and cognitive function (Buckner, 2013; Schmahmann, 2019) and has been implicated in psychiatric disorders such as schizophrenia, bipolar disorder, and autism (Phillips et al., 2015; Villanueva, 2012). Measuring cerebellar structures and analyzing potential brain changes in such disorders is dependent on accurate cerebellar segmentation. In the literature, various MRI-based segmentation methods are compared on different measures such as repeatability, reproducibility, and dice overlap to expert delineations, with varying results. (Carass et al., 2018; Sörös et al., 2021). In this study we compared the test-retest reliability of two different cerebellar segmentation methods, and determined the correlations between the segmented volumes.

Test-retest reliability for cerebellar volumes was calculated on an independent sample of n = 10 healthy volunteers acquired on two separate MRI scanners. T1-weighted structural images were acquired from n = 9 of the individuals (mean age = 35.76 years; range = [26.31-59.70]; 55% male) on a 3T GE Discovery MR750 scanner, and n = 9 of the individuals (mean age = 35.8 years; range = [25.24-60.52]; 55% male) on a 3T GE SIGNA Premier scanner. On each MRI scanner, individuals were scanned twice per session with a repositioning between scans, and a two-week interval between sessions, resulting in a total of four scans each. Cerebellar volumes were estimated with two different segmentation methods; the Sequence Adaptive Multimodal SEGmentation (SAMSEG; Puonti et al., 2016) and the Automatic Cerebellum Anatomical Parcellation using U-net with Locally Constrained Optimization (ACAPULCO; Han et al., 2020). Intra-class correlations (ICC) and 95% confidence intervals (CI) were computed for each combination of segmented volume, scan platform and segmentation tool, resulting in four measures from SAMSEG (grey- and white matter for each hemisphere) and an additional 28 measures of the cerebellar lobules from ACAPULCO. Furthermore, mean overall volumes and standard deviations were calculated, with a subsequent analysis of Pearson correlations to establish agreement between the methods.

We found ICCs > 0.9 for both methods between all sessions and scans within MRI scanner, indicating excellent reliability, apart from three lobules from ACAPULCO: Left I-III (MR750: ICC = 0.79; CI = 0.54-0.94, Premier: ICC = 0.8; CI = 0.57-0.94), Right I-III (MR750: ICC = 0.73; CI = 0.46-0.92, Premier: ICC = 0.85; CI = 0.66-0.96), and Vermis X (MR750: ICC = 0.89; CI = 0.73-0.97, Premier: ICC = 0.96; CI = 0.9-0.99). Correlations for overall volumetric output between the segmentation methods were high for whole-, hemispheric-, and grey matter-volumes (r = 0.96-0.97). However, correlations for white matter volumes were lower (r = 0.37-0.42).

Both methods showed high test-retest reliability for global cerebellum measures. The most notable difference was the estimation of white matter volumes, resulting in low correlation between approaches. The segmentation of white matter differs between SAMSEG and ACAPULCO, with the former including part of the pons and most white matter branches and the latter only isolating the white matter of the Corpus Medullare (Bogovic et al., 2013). Selecting an approach for volumetric analyses of the cerebellum should thus be made not only based on the reliability of the method, but also depending on the measure of interest.

Segmentation and Parcellation ¹

Cortical Anatomy and Brain Mapping

Neuroanatomy Other

Brain Atlases

Anatomical MRI ²

Cerebellum

Cortex

Segmentation

STRUCTURAL MRI

White Matter