Non-invasive visualization and segmentation of the dentate nucleus is helpful for characterizing neurological diseases. Therefore, we set up an automatic segmentation strategy relying on a convolutional neural network (CNN) for the delineation of the dentate nucleus based on quantitative susceptibility maps. We trained the network on 101 healthy controls and 118 patients suffering from various types of cerebellar ataxia. We were able to demonstrate that the CNN accurately segments the dentate nuclei in 26 healthy controls and 21 SCA6 patients with volume estimates being in agreement with literature.
Data Acquisition: 139 patients with cerebellar ataxia of different types (e.g., SCA1, SCA3, Friedreich ataxia), as well as 127 healthy controls, underwent MRI. Multi-echo gradient-echo imaging (TE1-4= 6.47ms/17.23ms/27.99ms/38.75ms, TR=62ms, flip angle [FA]=17°, bandwidth [BW]1-4=120Hz/px, voxel size=0.5mm×0.5mm×0.5mm) and T1-weighted imaging (MP-RAGE, voxel size=1mm×1mm×1mm, TE=3.26ms, TR=2530ms, inversion time=1100ms, FA=7°, BW=200Hz/px) were performed at 3T.
Data Processing: Quantitative susceptibility maps were calculated from the gradient-echo phase images using V-SHARP10, 11 and HEIDI.12 The dentate nuclei were manually traced in both hemispheres on the axial, sagittal and coronal susceptibility maps by at least one rater. For 47 subjects the dentate nuclei were delineated by two independent raters.
Segmentation: As a preprocessing step the cerebellum was identified based on the T1-weighted images using a dedicated CNN.13 The cerebellum mask was then transferred to the susceptibility maps via rigid-body registration and susceptibility values outside the cerebellum were set to 0. The signal intensities of the cerebellum-masked susceptibility maps were normalized across the study population using decile normalization.14 A two-pathway, 3D, 12 layer deep CNN was set up using the DeepMedic framework.15 The CNN was trained using the pre-processed susceptibility maps and the manual delineations of rater 1 of 219 participants (101 controls, 118 patients). The CNN relied on the Dice-Soerensen-Coefficient (DSC) as cost metric and the ADAM algorithm for optimization. The trained CNN was applied to 26 controls and 21 SCA6 patients.
Analysis: The CNN segmentation performance and the inter-rater variability were assessed using the DSC and the relative volume change with respect to the delineations of rater 1.
1. Tellmann S, Bludau S, Eickhoff S, Mohlberg H, Minnerop M, Amunts K. Cytoarchitectonic mapping of the human brain cerebellar nuclei in stereotaxic space and delineation of their co-activation patterns. Front Neuroanat. 2015;9:54.
2. Khadilkar S, Jaggi S, Patel B, Yadav R, Hanagandi P, Faria do Amaral LL. A practical approach to diseases affecting dentate nuclei. Clin Radiol. Jan 2016;71(1):107-119.
3. Haacke EM, Xu Y, Cheng YC, Reichenbach JR. Susceptibility weighted imaging (SWI). Magn Reson Med. Sep 2004;52(3):612-618.
4. Schweser F, Deistung A, Reichenbach JR. Foundations of MRI phase imaging and processing for Quantitative Susceptibility Mapping (QSM). Z Med Phys. Mar 2016;26(1):6-34.
5. Deistung A, Schweser F, Reichenbach JR. Overview of quantitative susceptibility mapping. NMR Biomed. Apr 2017;30(4).
6. Stefanescu MR, Dohnalek M, Maderwald S, et al. Structural and functional MRI abnormalities of cerebellar cortex and nuclei in SCA3, SCA6 and Friedreich's ataxia. Brain. May 2015;138(Pt 5):1182-1197.
7. Deistung A, Jäschke D, Göricke S, et al. Non-invasive Characterization of the Cerebellar Nuclei in Degenerative Cerebellar Ataxias (SCA3, SCA6) with Quantitative Susceptibility Mapping (QSM) at 3 Tesla. Paper presented at: Joint Annual Meeting ISMRM-ESMRMB2018; Paris, France.
8. Diedrichsen J, Maderwald S, Kuper M, et al. Imaging the deep cerebellar nuclei: a probabilistic atlas and normalization procedure. NeuroImage. Feb 1 2011;54(3):1786-1794.
9. Bazin P-L, Deistung A, Schäfer A, Turner R, Reichenbach J, Timmann D. Automated Segmentation of Cerebellar Nuclei from Ultra-High-Field Quantitative Susceptibility Maps with Multi-Atlas Shape Fusion. Paper presented at: Joint Annual Meeting ISMRM-ESMRMB2018; Paris, France.
10. Schweser F, Deistung A, Lehr BW, Reichenbach JR. Quantitative imaging of intrinsic magnetic tissue properties using MRI signal phase: an approach to in vivo brain iron metabolism? NeuroImage. Feb 14 2011;54(4):2789-2807.
11. Wu B, Li W, Guidon A, Liu C. Whole brain susceptibility mapping using compressed sensing. Magn Reson Med. Jan 2012;67(1):137-147.
12. Schweser F, Sommer K, Deistung A, Reichenbach JR. Quantitative susceptibility mapping for investigating subtle susceptibility variations in the human brain. NeuroImage. Sep 2012;62(3):2083-2100.
13. Jacobsen N, Deistung A, Timmann D, Reichenbach J, Güllmar D. Accurate Cerebellum segmentation using a 3D Convolutional Neural Network and fully connected CRF. Paper presented at: Joint Annual Meeting ISMRM-ESMRMB20182018; Paris, France.
14. Shah M, Xiao Y, Subbanna N, Francis S, Arnold DL, Collins DL, Arbel T. Evaluating intensity normalization on MRIs of human brain with multiple sclerosis. Medical Image Anal. 2011;15: 267–282.
15. Kamnitsas K, Ledig C, Newcombe VFJ, et al. Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation. Med Image Anal. Feb 2017;36:61-78.