Skip to main content
NCBI home page
Human Brain Mapping logoLink to Human Brain Mapping
. 2024 Aug 21;45(12):e26816. doi: 10.1002/hbm.26816

Pooled analysis of multiple sclerosis findings on multisite 7 Tesla MRI: Protocol and initial observations

Daniel M Harrison 1,2,✉, Seongjin Choi 1, Rohit Bakshi 3, Erin S Beck 4,5, Alexis M Callen 3, Renxin Chu 3, Jonadab Dos Santos Silva 4, Dumitru Fetco 6,7, Matthew Greenwald 5, Shannon Kolind 8, Sridar Narayanan 6,7, Serhat V Okar 5, Molly K Quattrucci 3, Daniel S Reich 5, David Rudko 6, Bretta Russell‐Schulz 8, Matthew K Schindler 9, Shahamat Tauhid 3, Anthony Traboulsee 8, Zachary Vavasour 8, Jonathan D Zurawski 3
PMCID: PMC11339124  PMID: 39169546

Abstract

Although 7 T MRI research has contributed much to our understanding of multiple sclerosis (MS) pathology, most prior data has come from small, single‐center studies with varying methods. In order to truly know if such findings have widespread applicability, multicenter methods and studies are needed. To address this, members of the North American Imaging in MS (NAIMS) Cooperative worked together to create a multicenter collaborative study of 7 T MRI in MS. In this manuscript, we describe the methods we have developed for the purpose of pooling together a large, retrospective dataset of 7 T MRIs acquired in multiple MS studies at five institutions. To date, this group has contributed five‐hundred and twenty‐eight 7 T MRI scans from 350 individuals with MS to a common data repository, with plans to continue to increase this sample size in the coming years. We have developed unified methods for image processing for data harmonization and lesion identification/segmentation. We report here our initial observations on intersite differences in acquisition, which includes site/device differences in brain coverage and image quality. We also report on the development of our methods and training of image evaluators, which resulted in median Dice Similarity Coefficients for trained raters' annotation of cortical and deep gray matter lesions, paramagnetic rim lesions, and meningeal enhancement between 0.73 and 0.82 compared to final consensus masks. We expect this publication to act as a resource for other investigators aiming to combine multicenter 7 T MRI datasets for the study of MS, in addition to providing a methodological reference for all future analysis projects to stem from the development of this dataset.

Keywords: 7 T, multicenter, multiple sclerosis

Harrison et al. describe methods for data pooling and analysis developed by team members from five institutions performing 7 T MRI research in multiple sclerosis. This manuscript will act as a reference for future analyses performed on a large 7‐T MRI dataset, which will eventually contain more than 1000 scans.

graphic file with name HBM-45-e26816-g008.jpg

Practitioner Points.

  • Large, multicenter studies may help confirm or refute conclusions drawn from prior, single‐center reports of the use of 7 T MRI for multiple sclerosis research.

  • Our group has developed a large, multicenter collaboration for 7 T MRI as applied to multiple sclerosis, along with associated methods for data sharing and analysis in a unified manner.

  • In addition to acting as a methodological reference for our future analyses to stem from this collaboration, we encourage future researchers to utilize the methods described here in similar projects going forward.

1. INTRODUCTION

Magnetic resonance imaging (MRI) is the main diagnostic and prognostic tool for patients with multiple sclerosis (MS). Visualization of MS pathology by MRI has brought about substantial advances in MS care, including earlier and more accurate diagnoses and the ability to monitor the effects of treatment. Despite these advances, it is well known that standard clinical MRI fails to visualize much of MS‐related pathology (Barkhof, 2002; Kutzelnigg et al., 2005). Although current/standard MRI performs well as a tool to measure acute inflammation and focal white matter (WM) demyelination in MS, it has been less successful as a tool to quantify gray matter (GM) demyelination, meningeal inflammation, and chronic active inflammation. For this reason, researchers continue to develop new imaging technologies—hoping to bring us closer to in vivo quantification of MS that mirrors histopathology.

The introduction of ultra‐high field, 7 T MRI to MS research has led to increased sensitivity to tissue damage versus conventional MRI and thus provides many important insights into MS pathophysiology. Cortical lesion (CL) burden, as visualized on 7 T MRI, is strongly related to disability, including an independent relationship between CLs and cognitive impairment (Harrison, Roy, et al., 2015; Nielsen et al., 2013). Lesions seen on 7 T MRI in deep GM structures, such as the thalamus, are also linked with MS‐related disability (Harrison, Oh, et al., 2015; Zurawski et al., 2020). Visualization of paramagnetic rims around WM lesions on 7 T susceptibility images has provided a means to identify chronic‐active lesions on MRI (Absinta et al., 2013; Bagnato et al., 2011; Harrison et al., 2016; Yao et al., 2012). Seven Tesla MRI also confirms the widespread nature of subpial cortical demyelination and demonstrates regions of blood‐cerebrospinal fluid (CSF) barrier breakdown (Cohen‐Adad et al., 2011; Harrison et al., 2017; Harrison et al., 2024; Jonas et al., 2018; Zurawski et al., 2020). To date, however, nearly all 7 T MRI research projects have been small, single‐center studies—limiting widespread applicability. Seven Tesla MRI research in MS has also been hampered by a paucity of widely available, validated image analysis tools. Seven Tesla image analysis has been mostly manual, which is time‐consuming, error‐prone, and infeasible for large scale applications.

Recognizing the need to accelerate translation of 7 T MRI for clinical use in MS, members of the North American Imaging in MS (NAIMS) Cooperative (Oh et al., 2018) formed the NAIMS 7 T MRI in MS Working Group. As part of this initiative, members of the Working Group launched a study that aims to standardize and characterize 7 T MRI correlates of pathology in MS through pooling of data from existing 7 T MRI studies at multiple sites. Our initial targets are cortical and deep GM lesions, paramagnetic rim lesions (PRLs), and meningeal contrast enhancement (MCE). In this manuscript, we describe the methods undertaken to gather and combine our multisite data, develop common processing methods to unify the data for collective use, and provide initial observations and lessons learned during this process. We intend for this publication to serve as a resource for future groups wishing to pursue similar collaborations and as a background reference for all future publications stemming from this endeavor.

2. MATERIALS AND METHODS

2.1. Data source

A multicenter collaborative was developed between the MS Centers of the University of Maryland, Baltimore (UMB), Brigham and Women's Hospital (BWH) at Harvard University and the University of Pennsylvania (UPenn), along with The Montreal Neurological Institute (MNI) at McGill University and the Translational Neuroradiology Section at the National Institute of Neurological Disorders and Stroke (NINDS). Each center pledged to contribute all qualifying 7 T MRI scans of the brain performed in previous and ongoing research using participants with MS. The potential number of subjects with imaging data available at each site at the time of drafting of this manuscript is described in Table 1. Studies performed at each contributing site were approved and overseen by local institutional review boards (IRB) and participants signed informed consent, including permission for sharing of anonymized or limited data set versions of their images/data.

TABLE 1.

Potential number of scans and study characteristics from each contributing site.

Site Number of unique subjects Total # of scans Range of follow‐up (years) Disability scales available at each visit
UMB 138 368 0–7.7 EDSS, MSFC, BICAMS, MFIS, BDI‐FS, LCVA
BWH 724 867 0–4.8 EDSS, T25FW, SDMT*, PASAT*
UPenn 36 55 0–1.0 EDSS
MNI 47 63 0–2.5 EDSS, PASAT*, SDMT*
NINDS 207 437 0–5.0 EDSS, MSFC, SDMT, LCVA
Total 1152 1790 0–7.7 ‐
Open in a new tab

Abbreviations: UMB, University of Maryland, Baltimore; BWH, Brigham and Women's Hospital; UPenn, University of Pennsylvania; MNI, Montreal Neurological Institute; NINDS, Translational Neuroradiology Section, National Institute of Neurological Disorders and Stroke; EDSS, Expanded Disability Status Scale; MSFC, Multiple Sclerosis Functional Composite; BICAMS, Brief International Cognitive Assessment for MS; MFIS, Multiple Sclerosis Fatigue Impact Scale; BDI‐FS, Beck Depression Inventory, Fast Screen; LCVA, Low Contrast Visual Acuity; T25FW, Timed 25 Foot Walk; SDMT, Symbol Digits Modalities Test; PASAT, Paced Auditory Serial Addition Test; *, available on a portion of the dataset.

2.2. MRI protocols

Acquisition protocols at each site are described in Table 2. As this is a retrospective study, MRI protocols were not prospectively unified and clear differences in hardware and acquisition are evident. Notably, data in this study are derived from images obtained on scanners by two manufacturers (Philips, Siemens), with various configurations and coil types.

TABLE 2.

MRI protocols and acquisition parameters.

Site Time epoch of acquisition Manufacturer/model Coil Sequence name Resolution (mm in each plane) 2D or 3D # of echoes TR (ms) TE (ms) TI (ms) FA (degrees) Acceleration type Acceleration factor GBCA Acquisition time (min:s)
BWH 2017‐Present Siemens/Magnetom Terra Volume transmit/32 receive FLAIR 0.7 × 0.7 × 0.7 3D 1 9000 301 2500 90 GRAPPA 3 +/− 8:08
NINDS 2015–present Siemens/Magnetom Volume transmit/32 receive FLAIR 0.8 × 0.8 × 0.8 3D 1 9000 271 2600 T2 var CAIPIRINHA 3 ‐ 7:14
UPenn 2019–2023 Siemens/Magnetom Terra Volume transmit/32 receive FLAIR 0.8 × 0.8 × 0.8 3D 1 6000 467 2400 T2 var GRAPPA 3 ‐ 7:59
MNI 2021–2022 Siemens/Magnetom Terra Volume transmit/32 receive FLAIR 0.7 × 0.7 × 0.7 3D 1 9000 301 2250 T2 var GRAPPA 3 +/− 8:08
UMB 2013–2018 Philips/Achieva Volume transmit/32 receive FLAIR 0.7 × 0.7 × 0.7 3D 1 8000 400 2077 90 SENSE 2 × 3 +/− 10:48
UMB 2019–present Philips/Achieva 8 channel transmit/32 receive FLAIR 0.5 × 0.488 × 0.488 3D 1 8000 300 2200 70 SENSE 2.5 × 3 +/− 9:04
BWH 2017–present Siemens/Magnetom Terra Volume transmit/32 receive GRE 0.8 × 0.8 × 0.8 2D 1 935 20 N/A 52 GRAPPA 4 ‐ 8:25
BWH 2022‐Present Siemens/Magnetom Terra Volume transmit/32 receive GRE 0.8 × 0.8 × 0.8 3D 6 44 6, 12, 17.99, 24, 29.99, 35.47 N/A 13 GRAPPA 2 ‐ 8:45
NINDS 2015–present Siemens/Magnetom Volume transmit/32 receive GRE 0.2 × 0.2 × 1.0 2D 2 1320 15, 32 N/A 50 GRAPPA 2 ‐ 8:46 per 25 mm slab
NINDS 2019–present Siemens/Magnetom Volume transmit/32 receive GRE 0.3 × 0.3 × 1.5 2D 2 2980 15, 35 N/A 60 GRAPPA 3 ‐ 8:37
NINDS 2020–present Siemens/Magnetom Volume transmit/32 receive GRE 0.5 × 0.5 × 0.5 3D 4 74 18, 29.5, 50.0, 52.4 N/A 10 GRAPPA/SENSE 3 ‐ 11:50 per 30 mm slab
UMB 2013–2018 Philips/Achieva Volume transmit/32 receive GRE 0.7 × 0.7 × 0.7 3D 5 28.87 5, 10, 15, 20, 25 N/A 10 SENSE 2.5 × 2 +/− 8:14
UMB 2019–present Philips/Achieva 8 channel transmit/32 receive GRE 0.66 × 0.66 × 0.7 3D 5 28 5, 10, 15, 20, 25 N/A 12 SENSE 2 × 2 ‐ 8:49
UPenn 2019–present Siemens/Magnetom Terra Volume transmit/32 receive GRE 0.8 × 0.8 × 0.8 3D 6 40 6, 12, 17.9 9, 24, 29.99, 35.47 N/A 13 GRAPPA 2 ‐ 7:38
MNI 2015–2019 Siemens/ Magnetom 8‐channel transmit/32 receive GRE 0.6 × 0.6 × 1.0 3D 6 40 6.12, 11.22, 16.32, 21.42, 26.52,31.62 N/A 13 GRAPPA 2 ‐ 6:39
NINDS 2015‐Present Siemens/Magnetom Volume transmit/32 receive MP2RAGE 0.7 × 0.7 × 0.7 3D 1 6000 3.02 800/2700 4/5 GRAPPA 3 ‐ 10:08
UMB 2013–2018 Philips/Achieva Volume transmit/32 receive MP2RAGE 0.7 × 0.688 × 0.688 3D 1 8250 (6.9 readout) 2.1 1000/3300 5/5 SENSE 2 × 2 +/− 9:20
UPenn 2019–present Siemens/Magnetom Terra Volume transmit/32 receive MP2RAGE 0.7 × 0.7 × 0.7 3D 1 5000 3.54 800/2700 4/5 GRAPPA 3 ‐ 9:55
MNI 2015–2019 Siemens/Magnetom 8‐channel transmit/32 receive MP2RAGE 0.7 × 0.7 × 0.7 3D 1 6000 2.73 800/2700 4/5 GRAPPA 3 ‐ 10:14
MNI 2021–2022 Siemens/Magnetom Terra Volume transmit/32 receive MP2RAGE 0.7 × 0.7 × 0.7 3D 1 4300 2.31 1000/3200 4/4 GRAPPA 3 +/− 10:16
UMB 2019–present Philips/Achieva 8 channel transmit/32 receive MP2RAGE 0.7 × 0.688 × 0.688 3D 1 8250 (6.9 readout) 2.2 1000/3300 5/5 SENSE 2 × 2 +/− 9:38
BWH 2017–present Siemens/Magnetom Terra Volume transmit/32 receive MP2RAGE 0.7 × 0.7 × 0.7 3D 1 4540 3.43 1040/3200 4/4 GRAPPA 4 +/− 8:07
UMB 2013–2018 Philips/Acheiva Volume transmit/32 receive MPRAGE 1.0 × 1.0 × 1.2 3D 1 4.7 2.1 N/A 7.0 SENSE 2 × 2.5 +/− 1:14
UMB 2019–present Philips/Achieva 8 channel transmit/32 receive MPRAGE 1.0 × 1.0 × 1.0 3D 1 5.0 1.81 N/A 7 SENSE 2 × 2 +/− 2:15
BWH 2017–present Siemens/Magnetom Terra Volume transmit/32 receive T2 SPACE 0.7 × 0.7 × 0.7 3D 1 2070 89 N/A T2 variable GRAPPA 3 ‐ 9:16
NINDS 2015–present Siemens/Magnetom Volume transmit/32 receive T2*w EPI 0.5 × 0.5 × 0.5 3D 1 52 23 N/A 10 EPI factor 15 ‐ 3:40 per 88 mm slab
Open in a new tab

Abbreviations: TR, repetition time; TE, echo time; TI, inversion time; FA, flip angle; GBCA, gadolinium‐based contrast agent (UMB, gadoteridol; BWH, gadoterate meglumine; MNI, gadobutrol); SENSE, SENSitivity Encoding; GRAPPA, GeneRalized Autocalibrating Partial Parallel Acquisition; MPRAGE, magnetization prepared rapid acquisition gradient echo; MP2RAGE, magnetization prepared 2 rapid acquisition gradient echo; FLAIR, fluid attenuated inversion recovery; GRE, gradient recall echo.

2.3. Data unification and harmonization

Differences in hardware and acquisition methods between studies posed challenges for common data analyses. To reduce this variability, we elected to utilize only sequences with similar acquisition methods and to perform image processing methods that would reduce variation without compromising sensitivity to pathology. Thus, not all subjects/scans/sequences available at each site were contributed to the common dataset.

Upon review of the protocols described in Table 2, we identified three acquisition methods with clinical/pathologic relevance to MS that were similar enough between sites for utilization in this study: magnetization prepared 2 rapid acquisition gradient echo (MP2RAGE), 3‐dimensional (3D) fluid attenuated inversion recovery (FLAIR), and multi‐echo, whole‐brain 3D gradient recall echo (ME‐GRE). Only these sequences from each site were utilized for image processing and data analysis. Our data processing methodology is summarized in Figure 1. The processing scripts necessary to replicate this methodology are available at https://github.com/s-j-choi/Pooled-7T-Data-Processing/tree/main.

FIGURE 1.

FIGURE 1

Open in a new tab

Data processing scheme. MP2RAGE, magnetization prepared 2 rapid acquisition gradient echo; Gd, post‐contrast; INV1, 1st inversion; INV2, 2nd inversion; T1‐w, T1‐weighted; T1‐w UNI, uniform T1‐weighted; denoised T1‐w, T1w UNI × INV2_N4 (N4, bias‐field corrected); FLAIR, fluid attenuated inversion recovery; GRE, gradient recalled echo; mag, magnitude; QSM, quantitative susceptibility mapping; LTS‐dFLAIR, LTS‐Gd‐FLAIR − FLAIR (LTS, least trimmed squares); prcnt‐dFLAIR, 100 × (LTS − dFLAIR)/FLAIR.

Although some site data had on‐scanner MP2RAGE processing outputs, to achieve uniformity in processing outputs, we elected to utilize the reconstructed acquisition files and perform offline processing of all images uniformly. A T1‐weighted (T1‐w) image was first generated using FSL/fslmaths, utilizing inversion 1 (INV1), and inversion 2 (INV2) images from the MP2RAGE acquisition, as previously described (Marques et al., 2010). The T1‐w image was multiplied by INV2 to suppress the background noise (Fujimoto et al., 2014), which was followed by N4 bias field correction (Tustison et al., 2010). This resulted in a denoised T1‐w image, which was used for skull stripping and co‐registration using ANTs (Avants et al., 2011).

The T1‐w image (prior to denoising) was used to calculate the T1‐map using a Matlab code provided by the developer (available at https://github.com/JosePMarques/MP2RAGE-relatedscripts). The denoised T1‐w image was linearly transformed to the MNI template, which had been resampled to an isotropic 0.7 mm resolution. The transformation matrix from this process was saved for future use.

If post‐contrast (Gd) MP2RAGE images were available, they were processed to create denoised‐Gd‐T1‐w and Gd‐T1‐map images by the same processes as above. The denoised Gd‐T1‐w image was first coregistered to the denoised‐T1‐w image, and then to the MNI template using the saved transformation matrix. Any images registered to the T1‐w images were transformed to MNI space using a single transformation matrix that combined the intra‐subject, intra‐visit image‐to‐T1w transformation with the T1w‐to‐MNI space transformation. Lastly, a dilated brain mask (“outskull mask”) of the denoised‐T1‐w was created using FSL/bet for subsequent use in processing FLAIR images.

FLAIR images first underwent N4 inhomogeneity correction followed by registration to the denoised T1‐w in its native space. The transformation matrix obtained from this registration was then combined with the transformation matrix for the registration of the T1‐w image to MNI space. This combined transformation was applied to the N4‐corrected FLAIR using a single transformation.

If a post‐contrast (Gd) FLAIR was available, N4 inhomogeneity correction was performed followed by registration to the native pre‐contrast FLAIR. The transformation matrix obtained from this registration was combined with the one from the registration of pre‐contrast FLAIR to MNI space. This combined transformation was applied to the N4‐corrected Gd‐FLAIR.

Post‐contrast FLAIR images were normalized to the pre‐contrast FLAIR to facilitate subtraction imaging. This was performed using a Least Trim Squares (LTS) normalization procedure (Rousseeuw & Van Driessen, 2006), with exclusion of outliers using custom Python code adapted from prior work for identification of contrast enhancement on T1‐w images (Elliott et al., 2019). Regions outside of the brain were included in this normalization step due to the need to analyze dura for MCE. After normalization, a FLAIR subtraction image was obtained by subtracting the pre‐contrast FLAIR from the LTS‐normalized Gd‐FLAIR. A FLAIR % difference map was also calculated using the aforementioned two images. In the case that only Gd‐FLAIR was available and pre‐contrast FLAIR was not, the Gd‐FLAIR was directly coregistered to its corresponding denoised T1‐w image. Following this, it was then transformed to the MNI space. No normalization or subtraction images were created.

Given varying echo time ranges for each ME‐GRE sequence in this study, we elected to only utilize echo times in common, which resulted in use of echoes in the 5–26 ms range. An average magnitude image was created for each, but with dropping of the first echo if shorter than 10 ms due to poor contrast in white/gray matter. This image was then registered to the N4 corrected INV2 in native space using rigid transformation. This transformation was combined with the transformation matrix of T1‐w to MNI space and applied to the full magnitude and average magnitude images. The phase images for all echoes in the 5–26 ms range was unwrapped using a Laplacian algorithm with a 2D Gaussian filter (https://github.com/blakedewey/phase_unwrap). An average (with dropping of echoes <10 ms) unwrapped filtered phase image was created in native space and both this and the individual echo images are transformed to MP2RAGE‐MNI space.

Quantitative Susceptibility Mapping (QSM) processing was performed in native space using the Johns Hopkins University/Kennedy Krieger Institute QSM Toolbox (https://github.com/xuli99/JHUKKI_QSM_Toolbox). The processing involves several steps: phase unwrapping based on the Laplacian method, removal of the background and mapping of frequencies, and inversion of the dipole using a modified SFCR method (Bao et al., 2016). The estimation of quantitative susceptibility values was achieved through internal referencing to the CSF region with an R2* value of 5 Hz or lower. The outcomes of this process were then transformed to the MP2RAGE‐MNI space. As part of internal referencing for quantitative susceptibility values, an R2* map was generated by Auto‐Regression on Linear Operations (ALRO) algorithm (Pei et al., 2015).

2.4. Data storage and sharing

This study utilizes the NAIMS Imaging Repository (NAIMS‐IR), a web‐platform (naims-ir.msmri.com) hosted by the MSMRI Research Group at the University of British Columbia, for data storage and sharing. This resource was commissioned by NAIMS to facilitate data sharing among NAIMS investigators and between NAIMS investigators and outside parties. Each participating site uploads image files labelled by local subject identification numbers through a secure HTTPS transfer. Uploaded files are automatically sorted by participant, scan, and sequence and displayed on the NAIMS‐IR's web interface for centralized tracking. Participants data are assigned a new subject identification number within the NAIMS‐IR, which remains consistent for each additional scan uploaded for that participant. All image headers are reformatted by the NAIMS‐IR to only contain this subject identification and remove local subject identification labels. Additional clinical information (e.g., disability score) is uploaded to the NAIMS‐IR through an automatically generated spreadsheet template and attached to the associated participants and scans. The NAIMS‐IR backend implements the robust LAMP stack (Linux, Apache, MySQL, PHP) secured by the latest cybersecurity software. Crash‐consistent backups are vaulted to a remote site every 24 h. This resource and the data described in this study are available to members of NAIMS and outside researchers through research collaborative agreements. Please visit the website listed above or contact the corresponding author for more information.

2.5. Quality assurance review

We have enacted two steps of quality assurance review in this project. The initial review is performed by the individual running the processing pipeline described in the previous section. All outputs are reviewed for successful completion of each processing step for each image before upload to the NAIMS‐IR. Each image rater analyzing specific lesion types (see below) performs a second level of quality assurance review. This consists of reviewing images pertinent to the lesion identification task at hand for the presence of appropriate sequences for the lesion identification task and for the quality of these images. Images with very poor quality (due to motion, other artifacts, etc.) are eliminated in this process. Any images identified as “borderline” during this review are then reviewed by the team assigned to that particular lesion type as a group and a final decision on inclusion of the images is made.

2.6. Image analysis

Three types of MS pathology were chosen for analysis on the 7 T images: GM lesions (cortical and deep GM), PRLs, and MCE. Team members were split into focus groups for each of these pathologic findings on imaging and protocols for identification and quantification were developed based on each focus group's prior experience and review of appropriate literature. A standard operating procedures (SOP) manual for image analysis was developed based on these discussions (see Data S1, Supporting Information).

The protocol for identification of cortical and deep GM lesions requires visualization of co‐registered T1‐w, FLAIR, and mean magnitude ME‐GRE images. Images are reviewed and annotated using ITK‐SNAP (Yushkevich et al., 2006) software. Examples of this process are shown in Figure 2 and definitions of the CL and deep GM lesion subtypes are described in Table 3. Each scan is to be masked by two independent reviewers. A union of these two independently annotated lesion masks are then generated and a third reviewer provides a consensus review to either keep or eliminate any lesions that were only identified by one of the two initial reviewers. This process was chosen due to initial observations that many lesions can be missed by each reviewer, and thus a full consensus review is necessary to maximize sensitivity while maintaining specificity. CLs are initially identified on the T1‐w MP2RAGE image as a hypointensity in the cortex, visually distinct from surrounding gray matter that is at least two adjacent voxels in one direction in size, and is confirmed in at least two orthogonal planes. CLs should not be identified if they have a linear/tubular appearance in any of the three orientations (suspicious for vessel or perivascular space) and should be carefully evaluated to avoid identifying as lesion an area of the cortex whose abnormal signal intensity is the result of partial volume averaging with CSF or plane of orientation of an image slice in relation to the cortex. All CLs identified on T1‐w must be confirmed with a corresponding hyperintensity either on the FLAIR or ME‐GRE image prior to masking. We also require that a minimum of approximately 1/4 of the lesion volume of a leukocortical CL be clearly within the GM to avoid inclusion of juxtacortical lesions. Due to the difficulty in accurately distinguishing between intracortical (histopathologic type II) and subpial (histopathologic type III/IV) (Kidd et al., 1999; Peterson et al., 2001), we categorized both lesion types in a combined category called “intracortical/subpial.”

FIGURE 2.

FIGURE 2

Open in a new tab

Cortical lesion masking example. Shown is a cortical lesion (yellow arrow) identified on T1‐w (a) and confirmed on FLAIR (b) and ME‐GRE (c). The lesion was then masked (d) as its borders appeared on T1‐w.

TABLE 3.

Cortical and deep GM lesion categories.

Lesion type Definition
Leukocortical Lesion traversing the WM/cortical GM boundary
Intracortical/subpial Lesion exclusive to cortical GM
Thalamic subependymal Lesion along ventricular surface of thalamus
Thalamic body Lesion in body of thalamus, not in contact with ventricular surface
Caudate subependymal Lesion along ventricular surface of caudate
Caudate body Lesion in body of caudate, not in contact with ventricular surface
Putamen Lesion within the putamen
Open in a new tab

Deep GM lesions are evaluated initially on T1‐w images and are defined as a hypointensity in the thalamus, caudate or putamen that is of a minimum size of at least two adjacent voxels in one plane, is present on at least two slices in at least one plane and is confirmed on at least two of three planes. Any lesion identified on T1‐w images must then be confirmed with a corresponding hyperintensity on FLAIR or ME‐GRE. Special attention is paid to avoid segmenting any perivascular spaces by ensuring that the hypointensity on T1‐w does not correspond to a hypointensity on FLAIR or ME‐GRE and that neither demonstrates a linear/tubular appearance. Special attention is also paid to avoid segmenting any areas of hypointensity on T1‐w images that correspond to natural boundaries between thalamic nuclei.

Review of images for the presence of PRLs is also performed using ITK‐SNAP. An example of the PRL identification process is shown in Figure 3. This process involves three independent reviewers. Any lesion identified by at least two of three reviewers is included in a final consensus mask, whereas any lesions only identified by just one of three reviewers is eliminated as a potential false positive. We have used the NAIMS consensus guidelines (Bagnato et al., 2024) for the identification of PRLs. Initial review for WMLs is conducted on the T1‐w image, although if a FLAIR image is available this also may be used for guidance on WML location and borders. Determination for PRL presence/absence is conducted on a lesion‐by‐lesion basis, independently on the mean, unwrapped, filtered phase image and the QSM image (each conducted at least 1 week apart for each scan to avoid bias). If a lesion is determined to meet NAIMS PRL criteria, the lesion is masked as it appears on the T1‐w image.

FIGURE 3.

FIGURE 3

Open in a new tab

PRL identification. The yellow arrow indicates a WML identified on T1‐w (a), which is confirmed to meet criteria as a PRL on PHS (b) and QSM (c). The lesion is masked as it appears on T1‐w (d), with the propagation of this mask to PHS (e) and QSM (f).

Review of images for the presence of MCE is performed using MIPAV (https://mipav.cit.nih.gov/index.php). This process involves three independent reviewers. An example of this is shown in Figure 4. Any lesion identified by at least two of three reviewers is kept for a final consensus mask, whereas any lesions identified by just one of three reviewers is eliminated as a potential false positive. Reviewers used pre‐contrast FLAIR, LTS‐normalized post‐contrast FLAIR, FLAIR subtraction, pre‐contrast T1‐w, and post‐contrast T1‐w images. The definition used for MCE in this review is any region of hyperintensity in the meningeal layers and/or meningeal spaces on post‐contrast FLAIR that is either not present on pre‐contrast FLAIR or is substantially more hyperintense on post‐contrast than pre‐contrast FLAIR. Reviewers initially identify candidate regions of MCE as hyperintensities on the FLAIR subtraction image. This region is then reviewed on the other sequences to determine anatomic location and confirm enhancement after administration of a gadolinium‐based contrast agent (GBCA). Special attention is paid to avoid segmenting any regions that are hyperintense on FLAIR images due to poor CSF signal suppression, areas obscured by motion, hyperintensities on subtraction images that are due to imperfect registration, or other artifacts. Confirmed regions of MCE are then masked via differing labels as one of five subtypes, defined in Table 4.

FIGURE 4.

FIGURE 4

Open in a new tab

MCE identification. Pre‐contrast FLAIR (a), post‐contrast FLAIR (b), and a FLAIR subtraction map (c) were reviewed for the presence of hyperintensities that met criteria for MCE. The yellow arrow indicates the location of a focus of subarachnoid spread/fill LME, which was then masked on the subtraction map (d).

TABLE 4.

Meningeal contrast enhancement (MCE) subtypes.

Enhancement type Definition
Nodular LME Small, discrete, often spherical nodule of enhancement adherent to pial surface or in subarachnoid space
Spread/fill LME Larger, amorphous areas of enhancement in the subarachnoid space, often following the contours of sulci/gyri and filling part or all of a sulcus
Paravenous Enhancement tracking along the space immediately outside the walls of large cortical veins. The vessel lumen should correspond to a flow void hypointensity on FLAIR and corresponding hyperintensity on T1‐w
Dural nodule Small, discrete, often spherical nodules of enhancement adherent to or within the dura
Parasinus Regions of enhancement along the space immediately outside the boundary of venous sinuses
Open in a new tab

Abbreviation: LME, leptomeningeal enhancement.

2.7. Training and certification of image raters

To ensure uniformity of methods for lesion annotation, each image rater must undergo a rigorous training and certification process as outlined in our SOP manual. The trainee first must work with their local site principal investigator (PI) to demonstrate proficiency in use of the necessary tools in the relevant image display software. They then work with their site PI to demonstrate proficiency in using the methods described in the SOP manual for lesion annotation specific to the lesion type they will be assigned to review. The SOP manual, and accompanying powerpoint presentations, contain information not only on lesion definitions, but also guidance on use of software for annotation, common methods for handling and visualization of the images, and uniform approaches to image windowing and contrast adjustment. Once this training is completed, they analyze a set of testing/training images that consist of representative scans from each site that have been masked by multiple expert raters via consensus. The correspondence between the rater's annotations and the expert/consensus mask is then analyzed for both the Dice Similarity Coefficient (DSC) and the lesion overlap rate (LOR), the latter of which is defined as the proportion of lesions overlapping between the two masks. Trainees must meet a minimum threshold of DSC 0.6 and LOR of 0.75 to pass certification.

3. RESULTS

Lesion masking for this project is underway and will be ongoing over the coming years. We will present results of our lesion quantification and relationships with clinical parameters in future publications. Here we describe quantification of the data set thus far, our initial assessments of the images being utilized in this study, and the impact of our training procedures.

3.1. Data set

At the time of writing this manuscript, 528 scans from 350 individuals have been uploaded to the NAIMS‐IR. Four‐hundred and seventy‐two scans have paired demographic data and 454 scans have paired clinical information and disability scores. Additional clinical data continues to be uploaded over time. The 528 scans include 4023 image series. Two‐hundred and eight‐eight scans have undergone processing, and 4634 processed series/image types are available for analysis.

Of the scans with paired demographic data available at the time of this writing, the median age is 47 years (range 18–70, interquartile range (IQR) 38–54). Three‐hundred and thirty‐nine participants are female and 133 are male. Of participants with paired clinical data, 243 participants have relapsing–remitting MS, 53 have secondary progressive MS, and 26 have primary progressive MS. Among these participants, the median EDSS score is 2.5 (range 0–7.5, IQR 1.5–3.5) and median symptom duration is 11.6 (range 0.1–47.8, IQR 6.0–9.8) years.

3.2. Image assessment

The acquisitions chosen for use in this retrospective, pooled data analysis study were selected to attain as much harmonization as possible despite the lack of a prospective, universal acquisition protocol. Further, our image processing methods were chosen to achieve further harmonization. Despite this, some differences remain that may impact future analyses and will need to be considered. Figure 5 shows representative images of denoised T1‐w MP2RAGE images from each of the five contributing sites. Despite similar acquisition parameters, gray/white contrast, field inhomogeneity, signal intensity properties, and brain coverage (particularly in the posterior fossa and anterior temporal lobes) vary between sites. These differences may impact the ability to quantify lesion burden in the gray matter and white matter.

FIGURE 5.

FIGURE 5

Open in a new tab

T1‐w comparison. Shown are the denoised T1‐w images from the MP2RAGE acquisition from UMB (a), NINDS (b), UPenn (c), MNI (d), and BWH (e). Differences in gray‐white contrast, brain coverage, and field inhomogeneity can be seen between sites.

Figure 6 similarly shows representative images of acquisitions depicting differences between 3D FLAIR images from four of the sites, with MNI not included due to the lack of this acquisition in their dataset. Differences in signal intensity profiles, gray/white contrast, and regions of signal dropout in the posterior fossa and anterior temporal lobes are even more pronounced with this image type. This may affect future analyses utilizing this image type, particularly for identification of GM lesions and MCE.

FIGURE 6.

FIGURE 6

Open in a new tab

FLAIR comparison. Shown are comparisons of the 3D FLAIR acquisition from UMB (a), NINDS (b), UPenn (c), and BWH (d). FLAIR not available from MNI. Note differences in tissue contrast, field inhomogeneity, and brain coverage between sites.

Figure 7 shows representative images of the mean magnitude 3D ME‐GRE from four sites, with NINDS not included due to the lack of this acquisition in their dataset. Brain coverage (both field of view and regions of signal loss), particularly in the posterior fossa and anterior temporal lobes are again quite different. These differences may impact our ability to perform lesion quantification with these images in the posterior fossa.

FIGURE 7.

FIGURE 7

Open in a new tab

ME‐GRE comparison. Shown are examples of ME‐GRE from UMB (a), UPenn (b), MNI (c), and BWH (d). ME‐GRE from NINDS not available. Note differences in tissue contrast, field inhomogeneity, and brain coverage between sites.

4. RESULTS OF TRAINING

The impact of unification of image analysis methods and training protocols were assessed at 3 time‐points, with DSC and LOR scores shown in Table 5. The first assessment was performed by expert raters before finalization of the SOP Manual and achievement of consensus on lesion definitions and the consensus mask set. DSC and LOR for this analysis represent comparisons between this initial lesion annotation and the final consensus masks. The second assessment was performed by new raters after receiving training by their site PI and review of the training materials. The third assessment was after receiving feedback on their attempts at lesion masking compared to the consensus masks and receiving further training. These results show marked improvement in both inter‐rater variability and overlap with the consensus masks with the use of both the SOP Manual and full training. The need for intense training and standardization was most evident for cortical and deep GM lesions, where both expert raters and raters at early stages of training showed marked variation in lesion identification.

TABLE 5.

Assessment of rater overlap with consensus masks.

Raters and timepoint Cortical and deep GM lesions PRL MCE
Median (range) DSC Median (range) LOR Median (range) DSC Median (range) LOR Median (range) DSC Median (range) LOR
Expert raters before finalization of SOP manual and consensus achievement 0.31 (0.05–0.50) 0.35 (0.03–0.64) 0.32 (0.00–0.72) 0.41 (0.00–0.73) 0.29 (0.15–0.31) 0.81 (0.66–0.84)
Trainee raters, after first round of training 0.17 (0.11–0.74) 0.27 (0.27–0.94) 0.77 (0.11–0.87) 0.86 (0.59–1.00) 0.46 (0.45–0.58) 0.75 (0.60–0.79)
Trainee raters, after additional training/feedback 0.73 (0.65–0.80) 0.95 (0.88–0.97) 0.82 (0.79–0.87) 1.00 (0.71–1.00) 0.80 (0.66–0.88) 1.00 (0.92–1.00)
Open in a new tab

Abbreviations: GM, gray matter; PRL, paramagnetic rim lesion; MCE, meningeal contrast enhancement; DSC, Dice Similarity Coefficient; LOR, lesion overlap rate; SOP, standard operating procedures.

5. DISCUSSION/CONCLUSION

This manuscript describes the process by which this working group is collecting and harmonizing multisite 7 T neuroimaging data in MS and developing unified methods for quantification of GM lesions, PRLs, and MCE on 7 T MRI. In the coming years, we expect to provide the results of ongoing work on this project, including inter‐rater reliability of our reading protocols and evaluations of the clinical relevance of each aspect of MS pathology visualized on 7 T MRI on a much larger scale than any prior reports. We also aim to utilize our work to facilitate development of automated methods for identification of MS pathology on 7 T MRI, which will further facilitate future 7 T MRI research in MS.

In addition to laying the groundwork for future research, we believe that our initial observations in this study point to the need for development of prospectively unified 7 T MRI acquisition protocols. The differences noted in acquisition details in Table 2 and in the appearance of images noted in Figures 5, 6, 7 may lead to unavoidable difficulties and biases in the data contributed by each site. Lesions outside of field of view may be missed and variations in SNR and field inhomogeneity that occur between sites and between subjects are likely to contribute to varying ability to discriminate lesion from non‐lesioned tissue. Our processing methods were chosen to limit these biases as much as possible, but only prospective unification of acquisition protocols and uniformity of hardware, including more widespread use of parallel transmit coils, may truly overcome this in the future. We will also explore the impact of site variations on the identification of MS pathology in future analysis, potentially providing evidence for the need for unified protocols. In future publications, we will note that the increased power due to the larger sample sizes in our analyses may be limited by acquisition variability. To the contrary, it is also possible that this variability may lead to more universal conclusions, resulting in findings that are applicable to MS regardless of acquisition protocol.

Other research initiatives have demonstrated that unification of 7 T MRI acquisition is possible. The German Ultrahigh Field Imaging (GUFI) research group has performed multiple “traveling head” studies at various 7 T MRI imaging centers in Germany and Austria (Voelker et al., 2016; Voelker et al., 2021). In these studies, Siemens 7 T scanners of various configurations underwent pre‐study system calibration measures based on B0 and B1 field mapping. This was followed by the acquisition of a uniform acquisition protocol that included MP2RAGE and ME‐GRE. Their studies demonstrate excellent reproducibility of the acquired anatomic images between sites. The UK7T Network also has performed similar harmonization work (Clarke et al., 2020; Rua et al., 2020). This study group consists of multiple 7 T imaging centers with representative Philips and Siemens systems in the United Kingdom. This group has shown that similar scanner calibration measures and additional unification of image reconstruction can also result in excellent reproducibility of acquisition protocols that include both anatomic (i.e., MP2RAGE, GRE, T2‐TSE) and functional MRI sequences. Neither of these initiatives, however, has validated these protocols for visualization of pathology, such as conditions like MS. We hope that the image analysis methods developed as a part of our retrospective pooled data analysis project will establish a foundation for unified methods of MS pathology quantification on 7 T MRI that may be employed in future 7 T studies, including in prospective studies with harmonized acquisition protocols. To facilitate this growth, we have placed our image processing scripts in the public domain and our SOP manual is available as Data S1. We hope that these resources will be adopted by future 7 T studies in MS. Further, we hope to grow the small consortium represented in this paper over time to include other members of NAIMS and centers participating in GUFI and the UK7T Network.

The data presented here also highlights the need for rigorous processes for training of manual image raters and achievement of consensus on identification of MS pathology on 7 T MRI. The expert raters, each of whom had extensive prior experience in identification of MS pathology on 7 T MRI, showed very poor DSC and LOR compared to the eventual consensus masks. This was particularly the case for cortical and deep GM lesions and PRLs. An initial round of training of staff for participating in this project was mostly sufficient for PRLs and MCE but was not for cortical and deep GM lesions. Variability in the ability of individual raters to identify these aspects of MS on 7 T MRI has been previously noted (Beck et al., 2018; Harrison, Roy, et al., 2015) and should be taken into account whenever reviewing the results of any manually annotated 7 T MRI project and particularly when comparing the data from analyses performed by different research groups. We propose that any future 7 T MRI projects adhere to rigid lesion definitions and a rigorous certification process for all of those participating in manual lesion identification. To facilitate this, we will make our SOP manual available with this publication (see Data S1) and our consensus training dataset can be made available upon request.

We propose that application of the methods described here can be beneficial for future and ongoing projects focusing on 7 T MRI in MS. Most 7 T MRI studies in MS to date have been single‐center studies with small sample sizes and varying methods used for lesion visualization and quantification, resulting in sometimes conflicting results and reduced widespread applicability of the findings. We hope that future publications from this research group will help to resolve discrepancies in the literature with a larger sample size and unified methods. This would lead to more comparable and reproducible results across single and/or multi centered 7 T studies.

Finally, we hope that this work can serve as a methodological framework for data sharing and harmonization extending beyond 7 T MRI research to future collaborative neuroimaging projects. Utilization of the NAIMS‐IR will provide not only the ability to share data between sites on this project, but also all other NAIMS Cooperative member sites and outside investigators. Indeed, additional collaborative projects utilizing this dataset through the NAIMS‐IR have already been initiated. Large scale collaborations such as this study may assist with addressing many of the remaining unanswered questions in MS research.

CONFLICT OF INTEREST STATEMENT

Dr. Harrison has received research funding from EMD‐Serono and Roche‐Genentech. Dr. Harrison has received consulting fees from TG Therapeutics and EMD‐Serono and receives royalties from Up To Date, Inc. Dr. Kolind has received research support from Roche‐Genentech, Sanofi‐Genzyme and Biogen. Dr. Beck has received consulting fees from EMD‐Serono. Dr. Reich has received research funding from Abata and Sanofi. Dr. Bakshi has received speaking honoraria from EMD Serono and research support from Bristol Myers Squibb, EMD Serono, and Novartis. Dr. Narayanan has received research funding from F. Hoffman LaRoche and Immunotec, is a consultant for Sana Biotech and is a part‐time employee of NeuroRx Research. Dr. Traboulsee has received consulting fees from Roche‐Genentech, Sanofi‐Genzyme, Serono, Biogen and has research funding from Roche‐Genentech and Sanofi. Dr. Fetco is an employee of NeuroRx Research. Dr. Zurawski has received research support from the Elizabeth A. Kremer Research Foundation, Novartis Pharmaceuticals, NIH, I‐Mab Biopharma and the Race to Erase MS Foundation. Drs. Choi, Rudko, Schindler, and Tauhid have nothing to disclose. Ms. Callen and Quatrucci and Mr. Greenwald have nothing to disclose.

Supporting information

Data S1. Supporting Information.

HBM-45-e26816-s001.docx (15.2MB, docx)

ACKNOWLEDGEMENTS

We would like to thank Zahra Karimaghaloo, Colm Elliott, and Vladimir Fonov from NeuroRx Research for providing code to assist with LTS normalization. We also would like to thank others who have contributed to this project in the past, including Adelia Adelia, Douglas Arnold, Duha Awad, Luyun Chen, Youmna Jalk, Samar Khalil, Abigail Manning, and Madeline Seitz. We would also like to thank the MRI technologists and MRI physicists at the Kennedy Krieger Institute at Johns Hopkins University in Baltimore, Maryland, for assistance with acquisition of the 7 T MRI scans utilized by UMB in this project. We also would like to thank the Steering Committee of the NAIMS Cooperative for their support in this project. This study is funded by the National Institute of Neurological Disorders and Stroke (NINDS) grant R01NS122980 and intramural funding from the NINDS (Reich lab). The funders of this study had no role in study design, implementation, or manuscript preparation. Acquisition of 7 T MRIs from UMB were acquired under NINDS R01NS104403 (PI: Harrison), NINDS K23NS072366 (PI: Harrison), the Grant for Multiple Sclerosis Innovation (GMSI) program through EMD‐Serono, and an investigator‐initiated grant from Roche‐Genentech. Acquisition of 7 T MRIs from BWH were acquired under routine clinical care, and grants from the Race to Erase MS, Novartis, EMD Serono, and Bristol Myers Squibb. Acquisition of 7 T MRIs from UPenn were acquired under in‐kind funding through the Center for Neuroinflammation and Experimental Therapeutics at the University of Pennsylvania. Acquisition of 7 T MRIs from MNI were acquired under Canadian Institutes of Health Research FRN 84367 and PJT153005. Acquisition of 7 T MRIs at NINDS was funded by the NIH Intramural Research Program (ZIANS003119). SVO is supported by the National MS Society (Post‐doctoral Fellowship Grant, FG‐2208‐40289). Dr. Beck is supported by a Career Transition Award from the National MS Society (TA‐2109‐38412). The NAIMS‐IR was funded by a grant from the Conrad N. Hilton Foundation (Grant number 20220).

Harrison, D. M. , Choi, S. , Bakshi, R. , Beck, E. S. , Callen, A. M. , Chu, R. , Silva, J. D. S. , Fetco, D. , Greenwald, M. , Kolind, S. , Narayanan, S. , Okar, S. V. , Quattrucci, M. K. , Reich, D. S. , Rudko, D. , Russell‐Schulz, B. , Schindler, M. K. , Tauhid, S. , Traboulsee, A. , … Zurawski, J. D. (2024). Pooled analysis of multiple sclerosis findings on multisite 7 Tesla MRI: Protocol and initial observations. Human Brain Mapping, 45(12), e26816. 10.1002/hbm.26816

DATA AVAILABILITY STATEMENT

The authors of this publication will make the consensus training sets described in the Methods section available upon request. Please email the corresponding author with this request and an internet link will be provided for transfer. Due to local privacy regulations, access to the full dataset will require establishment of a data sharing agreement between any requestor and the NAIMS‐IR and will also require approval of each of the institutions contributing data to this project in the NAIMS‐IR. We welcome others to place requests for access and/or to contribute images of their own. Please contact the corresponding author for more details.

REFERENCES

  1. Absinta, M. , Sati, P. , Gaitan, M. I. , et al. (2013). Seven‐tesla phase imaging of acute multiple sclerosis lesions: A new window into the inflammatory process. Annals of Neurology, 74, 669–678. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Avants, B. B. , Tustison, N. J. , Song, G. , Cook, P. A. , Klein, A. , & Gee, J. C. (2011). A reproducible evaluation of ANTs similarity metric performance in brain image registration. NeuroImage, 54, 2033–2044. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Bagnato, F. , Hametner, S. , Yao, B. , van Gelderen, P. , Merkle, H. , Cantor, F. K. , Lassmann, H. , & Duyn, J. H. (2011). Tracking iron in multiple sclerosis: A combined imaging and histopathological study at 7 Tesla. Brain, 134, 3602–3615. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Bagnato, F. , Sati, P. , Hemond, C. C. , Elliott, C. , Gauthier, S. A. , Harrison, D. M. , Mainero, C. , Oh, J. , Pitt, D. , Shinohara, R. T. , Smith, S. A. , Trapp, B. , Azevedo, C. J. , Calabresi, P. A. , Henry, R. G. , Laule, C. , Ontaneda, D. , Rooney, W. D. , Sicotte, N. L. , … Absinta, M. (2024). Imaging chronic active lesions in multiple sclerosis: A consensus statement. Brain, 2024, awae013. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Bao, L. , Li, X. , Cai, C. , Chen, Z. , & van Zijl, P. C. (2016). Quantitative susceptibility mapping using structural feature based collaborative reconstruction (SFCR) in the human brain. IEEE Transactions on Medical Imaging, 35, 2040–2050. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Barkhof, F. (2002). The clinico‐radiological paradox in multiple sclerosis revisited. Current Opinion in Neurology, 15, 239–245. [DOI] [PubMed] [Google Scholar]
  7. Beck, E. S. , Sati, P. , Sethi, V. , Kober, T. , Dewey, B. , Bhargava, P. , Nair, G. , Cortese, I. C. , & Reich, D. S. (2018). Improved visualization of cortical lesions in multiple sclerosis using 7T MP2RAGE. American Journal of Neuroradiology, 39, 459–466. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Clarke, W. T. , Mougin, O. , Driver, I. D. , Rua, C. , Morgan, A. T. , Asghar, M. , Clare, S. , Francis, S. , Wise, R. G. , Rodgers, C. T. , Carpenter, A. , Muir, K. , & Bowtell, R. (2020). Multi‐site harmonization of 7 Tesla MRI neuroimaging protocols. NeuroImage, 206, 116335. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Cohen‐Adad, J. , Benner, T. , Greve, D. , Kinkel, R. P. , Radding, A. , Fischl, B. , Rosen, B. R. , & Mainero, C. (2011). In vivo evidence of disseminated subpial T2* signal changes in multiple sclerosis at 7 T: A surface‐based analysis. NeuroImage, 57, 55–62. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Elliott, C. , Wolinsky, J. S. , Hauser, S. L. , Kappos, L. , Barkhof, F. , Bernasconi, C. , Wei, W. , Belachew, S. , & Arnold, D. L. (2019). Slowly expanding/evolving lesions as a magnetic resonance imaging marker of chronic active multiple sclerosis lesions. Multiple Sclerosis, 25, 1915–1925. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Fujimoto, K. , Polimeni, J. R. , van der Kouwe, A. J. , et al. (2014). Quantitative comparison of cortical surface reconstructions from MP2RAGE and multi‐echo MPRAGE data at 3 and 7 T. NeuroImage, 90, 60–73. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Harrison, D. M. , Allette, Y. M. , Zeng, Y. , Cohen, A. , Dahal, S. , Choi, S. , Zhuo, J. , & Hua, J. (2024). Meningeal contrast enhancement in multiple sclerosis: Assessment of field strength, acquisition delay, and clinical relevance. PLoS One, 19, e0300298. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Harrison, D. M. , Li, X. , Liu, H. , Jones, C. K. , Caffo, B. , Calabresi, P. A. , & van Zijl, P. (2016). Lesion heterogeneity on high‐field susceptibility MRI is associated with multiple sclerosis severity. American Journal of Neuroradiology, 37, 1447–1453. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Harrison, D. M. , Oh, J. , Roy, S. , Wood, E. T. , Whetstone, A. , Seigo, M. A. , Jones, C. K. , Pham, D. , van Zijl, P. , Reich, D. S. , & Calabresi, P. A. (2015). Thalamic lesions in multiple sclerosis by 7T MRI: Clinical implications and relationship to cortical pathology. Multiple Sclerosis, 21, 1139–1150. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Harrison, D. M. , Roy, S. , Oh, J. , Izbudak, I. , Pham, D. , Courtney, S. , Caffo, B. , Jones, C. K. , van Zijl, P. , & Calabresi, P. A. (2015). Association of cortical lesion burden on 7‐T magnetic resonance imaging with cognition and disability in multiple sclerosis. JAMA Neurology, 72, 1004–1012. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Harrison, D. M. , Wang, K. Y. , Fiol, J. , Naunton, K. , Royal, W., III , Hua, J. , & Izbudak, I. (2017). Leptomeningeal enhancement at 7T in multiple sclerosis: Frequency, morphology, and relationship to cortical volume. Journal of Neuroimaging, 27, 461–468. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Jonas, S. N. , Izbudak, I. , Frazier, A. A. , & Harrison, D. M. (2018). Longitudinal persistence of meningeal enhancement on Postcontrast 7T 3D‐FLAIR MRI in multiple sclerosis. American Journal of Neuroradiology, 39, 1799–1805. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Kidd, D. , Barkhof, F. , McConnell, R. , Algra, P. R. , Allen, I. V. , & Revesz, T. (1999). Cortical lesions in multiple sclerosis. Brain, 122, 17–26. [DOI] [PubMed] [Google Scholar]
  19. Kutzelnigg, A. , Lucchinetti, C. F. , Stadelmann, C. , Brück, W. , Rauschka, H. , Bergmann, M. , Schmidbauer, M. , Parisi, J. E. , & Lassmann, H. (2005). Cortical demyelination and diffuse white matter injury in multiple sclerosis. Brain, 128, 2705–2712. [DOI] [PubMed] [Google Scholar]
  20. Marques, J. P. , Kober, T. , Krueger, G. , van der Zwaag, W. , Van de Moortele, P. F. , & Gruetter, R. (2010). MP2RAGE, a self bias‐field corrected sequence for improved segmentation and T1‐mapping at high field. NeuroImage, 49, 1271–1281. [DOI] [PubMed] [Google Scholar]
  21. Nielsen, A. S. , Kinkel, R. P. , Madigan, N. , Tinelli, E. , Benner, T. , & Mainero, C. (2013). Contribution of cortical lesion subtypes at 7T MRI to physical and cognitive performance in MS. Neurology, 81, 641–649. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Oh, J. , Bakshi, R. , Calabresi, P. A. , Crainiceanu, C. , Henry, R. G. , Nair, G. , Papinutto, N. , Constable, R. T. , Reich, D. S. , Pelletier, D. , Rooney, W. , Schwartz, D. , Tagge, I. , Shinohara, R. T. , Simon, J. H. , Sicotte, N. L. , & NAIMS Cooperative Steering Committee . (2018). The NAIMS cooperative pilot project: Design, implementation and future directions. Multiple Sclerosis, 24, 1770–1772. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Pei, M. , Nguyen, T. D. , Thimmappa, N. D. , Salustri, C. , Dong, F. , Cooper, M. A. , Li, J. , Prince, M. R. , & Wang, Y. (2015). Algorithm for fast monoexponential fitting based on auto‐regression on linear operations (ARLO) of data. Magnetic Resonance in Medicine, 73, 843–850. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Peterson, J. W. , Bo, L. , Mork, S. , Chang, A. , & Trapp, B. D. (2001). Transected neurites, apoptotic neurons, and reduced inflammation in cortical multiple sclerosis lesions. Annals of Neurology, 50, 389–400. [DOI] [PubMed] [Google Scholar]
  25. Rousseeuw, P. J. , & Van Driessen, K. (2006). Computing LTS regression for large data sets. Data Mining and Knowledge Discovery, 12, 29–45. [Google Scholar]
  26. Rua, C. , Clarke, W. T. , Driver, I. D. , Mougin, O. , Morgan, A. T. , Clare, S. , Francis, S. , Muir, K. W. , Wise, R. G. , Carpenter, T. A. , Williams, G. B. , Rowe, J. B. , Bowtell, R. , & Rodgers, C. T. (2020). Multi‐centre, multi‐vendor reproducibility of 7T QSM and R(2)* in the human brain: Results from the UK7T study. NeuroImage, 223, 117358. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Tustison, N. J. , Avants, B. B. , Cook, P. A. , Zheng, Y., Egan, A. , Yushkevich, P. A. , & Gee, J. C. (2010). N4ITK: Improved N3 bias correction. IEEE Transactions on Medical Imaging, 29, 1310–1320. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Voelker, M. N. , Kraff, O. , Brenner, D. , Wollrab, A. , Weinberger, O. , Berger, M. C. , Robinson, S. , Bogner, W. , Wiggins, C. , Trampel, R. , Stöcker, T. , Niendorf, T. , Quick, H. H. , Norris, D. G. , Ladd, M. E. , & Speck, O. (2016). The traveling heads: Multicenter brain imaging at 7 Tesla. Magma, 29, 399–415. [DOI] [PubMed] [Google Scholar]
  29. Voelker, M. N. , Kraff, O. , Goerke, S. , Laun, F. B. , Hanspach, J. , Pine, K. J. , Ehses, P. , Zaiss, M. , Liebert, A. , Straub, S. , Eckstein, K. , Robinson, S. , Nagel, A. N. , Stefanescu, M. R. , Wollrab, A. , Klix, S. , Felder, J. , Hock, M. , Bosch, D. , … Quick, H. H. (2021). The traveling heads 2.0: Multicenter reproducibility of quantitative imaging methods at 7 Tesla. NeuroImage, 232, 117910. [DOI] [PubMed] [Google Scholar]
  30. Yao, B. , Bagnato, F. , Matsuura, E. , Merkle, H. , van Gelderen, P. , Cantor, F. K. , & Duyn, J. H. (2012). Chronic multiple sclerosis lesions: Characterization with high‐field‐strength MR imaging. Radiology, 262, 206–215. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Yushkevich, P. A. , Piven, J. , Hazlett, H. C. , Smith, R. G. , Ho, S. , Gee, J. C. , & Gerig, G. (2006). User‐guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability. NeuroImage, 31, 1116–1128. [DOI] [PubMed] [Google Scholar]
  32. Zurawski, J. , Tauhid, S. , Chu, R. , Khalid, F. , Healy, B. C. , Weiner, H. L. , & Bakshi, R. (2020). 7T MRI cerebral leptomeningeal enhancement is common in relapsing‐remitting multiple sclerosis and is associated with cortical and thalamic lesions. Multiple Sclerosis, 26, 177–187. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Data S1. Supporting Information.

HBM-45-e26816-s001.docx (15.2MB, docx)

Data Availability Statement

The authors of this publication will make the consensus training sets described in the Methods section available upon request. Please email the corresponding author with this request and an internet link will be provided for transfer. Due to local privacy regulations, access to the full dataset will require establishment of a data sharing agreement between any requestor and the NAIMS‐IR and will also require approval of each of the institutions contributing data to this project in the NAIMS‐IR. We welcome others to place requests for access and/or to contribute images of their own. Please contact the corresponding author for more details.

Articles from Human Brain Mapping are provided here courtesy of Wiley

RESOURCES