Portable ultra-low-field brain MRI: test-retest reliability and correspondence to high-field MRI
Poster No:
2280
Submission Type:
Abstract Submission
Authors:
František Váša1, Carly Bennalick1, Niall Bourke1, Francesco Padormo2, Paul Cawley1, Tomoki Arichi1, Tobias Wood1, David Lythgoe1, Flavio Dell’Acqua1, Levente Baljer1, Sean Deoni3, Ashwin Venkataraman1, Rosalyn Moran1, Robert Leech1, Steven Williams1
Institutions:
1King's College London, London, United Kingdom, 2Hyperfine Inc., Guildford, CT, 3Bill and Melinda Gates Foundation, Seattle, WA
First Author:
Co-Author(s):
Introduction:
Ultra-low-field magnetic resonance imaging (MRI) scanners, such as the 64mT Hyperfine Swoop, promise to revolutionize neuroimaging [1]. The Hyperfine scanner runs using a standard electrical socket and is portable, enabling scanning at the bedside and in low- and middle-income countries with limited access to MRI. However, it is unclear whether 64mT scans can be used to reliably quantify tissue volume, and to what extent such measurements correspond to 3T high-field MRI.
Methods:
We recruited 23 healthy adult participants, with 2/3 male and 2/3 female participants in each of five strata: 20–29, 30-39, 40-49, 50-59 and 60-69 years old. Participants were scanned on a 3T MRI scanner (GE Premier) and two identical portable 64mT MRI scanners (Hyperfine Swoop) at different sites, using T1w and T2w scans. 3T scans were acquired at 1x1x1 mm resolution, while 64mT scans were acquired using both non-isotropic product sequences (T1w: 1.6x1.6x5 mm / T2w: 1.5x1.5x5 mm, with high resolution in the axial, sagittal or coronal plane), and a custom isotropic sequence (2.3x2.3x2.3 mm). We used multi-resolution registration (MRR) to super-resolve the three orthogonal non-isotropic 64mT scans into a single higher-resolution (T1w: 1.6x1.6x1.6 mm / T2w: 1.5x1.5x1.5 mm) scan [2] (Fig. 1).
We then used SynthSeg+ [3] to segment each scan into 98 structures and estimate their volumes. We assessed test-retest reliability of volume estimates using the intraclass correlation coefficient ICC(3,1) [4], hereafter referred to as ICC. We assessed correspondence of volumes from 64mT and 3T scans using both Pearson's r – to quantify linear correlation; and Lin's Concordance Correlation Coefficient (CCC) – to quantify exact agreement. Finally, to test whether 64mT reliability and correspondence to 3T is higher for larger regions, we quantified the association of regional ICC, Pearson's r and Lin's CCC to the average 3T volume of each region, using Spearman's ρ.
We repeated analyses on both T1w and T2w scans, across 64mT resolutions.
We then used SynthSeg+ [3] to segment each scan into 98 structures and estimate their volumes. We assessed test-retest reliability of volume estimates using the intraclass correlation coefficient ICC(3,1) [4], hereafter referred to as ICC. We assessed correspondence of volumes from 64mT and 3T scans using both Pearson's r – to quantify linear correlation; and Lin's Concordance Correlation Coefficient (CCC) – to quantify exact agreement. Finally, to test whether 64mT reliability and correspondence to 3T is higher for larger regions, we quantified the association of regional ICC, Pearson's r and Lin's CCC to the average 3T volume of each region, using Spearman's ρ.
We repeated analyses on both T1w and T2w scans, across 64mT resolutions.
Results:
We report summary statistics as Median [1st, 3rd quartile] across 98 segmentation labels (regions), or Md [Q1,Q3] (Fig. 2).
64mT scans showed high between-scanner test-retest reliability, across contrasts and scan resolutions. The most reliable volumes were yielded by T2w 64mT scans super-resolved using MRR (ICC Md [Q1,Q3] = 0.97 [0.95,0.99]; Fig. 2B). Reliability was similarly high for volumes extracted from MRR-super-resolved 64mT T1w scans, followed by marginally lower reliability for 64mT T2w and T1w scans at lower resolutions.
Volumes extracted from 64mT scans showed excellent correspondence to 3T across participants (as quantified by Pearson's r) but also slight systematic offsets in volume measurements (as quantified by Lin's CCC), with a tendency of 64mT MRI to underestimate 3T volumes. The highest correspondence to 3T was shown by 64mT T2w scans super-resolved using MRR, including both high correlation (Pearson's r Md [Q1,Q3] = 0.96 [0.93,0.98]) and agreement (Lin's CCC Md [Q1,Q3] = 0.90 [0.81,0.95]) (Fig. 2B). MRR-super-resolved 64mT T1w scans showed marginally lower linear correlation to 3T data than T2w scans, as well as lower agreement, followed by 2.3mm isotropic 64mT T2w and T1w scans, and non-isotropic scans.
Finally, both 64mT reliability and correspondence to 3T MRI tended to be higher for larger regions. Average regional (3T) volume was significantly associated to regional ICC (range of ρ across 2 contrasts x 5 resolutions; ρ = 0.42-0.55), as well as to Pearson's r (ρ = 0.32-0.52) and Lin's CCC (ρ = 0.22-0.47).
64mT scans showed high between-scanner test-retest reliability, across contrasts and scan resolutions. The most reliable volumes were yielded by T2w 64mT scans super-resolved using MRR (ICC Md [Q1,Q3] = 0.97 [0.95,0.99]; Fig. 2B). Reliability was similarly high for volumes extracted from MRR-super-resolved 64mT T1w scans, followed by marginally lower reliability for 64mT T2w and T1w scans at lower resolutions.
Volumes extracted from 64mT scans showed excellent correspondence to 3T across participants (as quantified by Pearson's r) but also slight systematic offsets in volume measurements (as quantified by Lin's CCC), with a tendency of 64mT MRI to underestimate 3T volumes. The highest correspondence to 3T was shown by 64mT T2w scans super-resolved using MRR, including both high correlation (Pearson's r Md [Q1,Q3] = 0.96 [0.93,0.98]) and agreement (Lin's CCC Md [Q1,Q3] = 0.90 [0.81,0.95]) (Fig. 2B). MRR-super-resolved 64mT T1w scans showed marginally lower linear correlation to 3T data than T2w scans, as well as lower agreement, followed by 2.3mm isotropic 64mT T2w and T1w scans, and non-isotropic scans.
Finally, both 64mT reliability and correspondence to 3T MRI tended to be higher for larger regions. Average regional (3T) volume was significantly associated to regional ICC (range of ρ across 2 contrasts x 5 resolutions; ρ = 0.42-0.55), as well as to Pearson's r (ρ = 0.32-0.52) and Lin's CCC (ρ = 0.22-0.47).
Conclusions:
Volume estimates from portable ultra-low-field MRI scans show excellent test-retest reliability and correspondence to high-field counterparts. Reliability and correspondence to 3T were highest for T2w 64mT scans super-resolved using MRR [2]. Our results pave the way for the modelling of individual deviations from the norm [5,6] and estimation of biomarkers such as brain age [7], across development [8] and disease [9].
Modeling and Analysis Methods:
Segmentation and Parcellation
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Cortical Anatomy and Brain Mapping 2
Subcortical Structures
Novel Imaging Acquisition Methods:
Anatomical MRI 1
Keywords:
Cortex
Segmentation
STRUCTURAL MRI
Other - Ultra Low Field MRI; Low Field MRI; Portable MRI; 64mT MRI; Reliability
1|2Indicates the priority used for review
Provide references using author date format
[2] Deoni S.C.L., et al. (2022). Simultaneous high-resolution T2 -weighted imaging and quantitative T2 mapping at low magnetic field strengths using a multiple TE and multi-orientation acquisition approach. Magnetic Resonance in Medicine, 88(3), 1273–1281.
[3] Billot B., et al. (2023). Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets. Proceedings of the National Academy of Sciences, 120(9), e2216399120.
[4] Chen G., et al. (2018). Intraclass correlation: Improved modeling approaches and applications for neuroimaging. Human Brain Mapping, 39(3), 1187–1206.
[5] Bethlehem R.A.I.*, Seidlitz J.*, White S.R.*, et al. (2022). Brain charts for the human lifespan. Nature, 604(7906), Article 7906.
[6] Rutherford S*, et al. (2022). The normative modeling framework for computational psychiatry. Nature Protocols, 17(7), Article 7.
[7] Cole J.H. et al. (2017). Predicting Age Using Neuroimaging: Innovative Brain Ageing Biomarkers. Trends in Neurosciences, 40(12), 681–690.
[8] Deoni S.C.L., et al. (2021). Accessible pediatric neuroimaging using a low field strength MRI scanner. NeuroImage, 238, 118273.
[9] Venkataraman A., et al. (2023). Automated brain image analysis in dementia using artificial intelligence: A roadmap for the development of clinical tools. OSF Preprints.