Precision brain morphometry using cluster scanning

Structural MRI is widely used to quantify brain morphometry in the study of development, aging, psychopathology, and neurodegeneration (i.e., regional brain volume and cortical thickness). For example, morphometric studies have found age-related cortical thinning and volumetric atrophy (Bethlehem et al., 2022; Fjell et al., 2014; Frangou et al., 2022; Raz et al., 2005; Salat et al., 2004; Sowell et al., 2003; Storsve et al., 2014), changes in the hippocampus in response to extensive training (Maguire et al., 2000), and accelerated brain atrophy in cognitive decline and neurodegenerative disease (Cox et al., 2021; Dickerson et al., 2009; Jack et al., 2018; Johnson et al., 2012; Keret et al., 2021). However, the utility of morphometric studies is limited by measurement error. Commonly used MRI-derived morphometric estimates have measurement errors of ~2–5% (e.g., Tustison et al., 2014). Measurement error limits statistical power and impacts sample size, longitudinal follow-up duration, and study cost.

Recent advances in MRI acceleration suggest a path toward higher precision morphometrics. Compressed sensing (CS) and methods based on wave-controlled aliasing in parallel imaging (Wave-CAIPI) allow for a rapid high-resolution T₁-weighted magnetization-prepared rapid gradient echo (MPRAGE) scan to be acquired in about a minute, approximately 1/5th the time of an MPRAGE with standard parallel imaging acceleration (Bilgic et al., 2015; Dieckmeyer et al., 2021; Mussard et al., 2020; Polak et al., 2018). Our recent explorations found that a single rapid scan collected with CS acceleration (CS; 1’12”) or a Wave-CAIPI acceleration (WAVE; 1’09”) both produce morphometric estimates with high test-retest reliability, high convergent validity, and an absolute measurement error similar to a longer T₁-weighted MPRAGE based on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) protocol (Elliott et al., 2023). These findings raise the possibility that multiple fast scans acquired in rapid succession could be used together to drive down measurement error beyond what is possible from a single traditional scan (see also Nielsen et al., 2019). If successful, this approach also promises to allow for greater study design flexibility, mitigate participant burden, and reduce the negative impacts of head motion because shorter scans allow less time for head motion and can be easily repeated when motion corrupted.

An established method for increasing precision is to pool estimates by taking the average of multiple independent measurements. Morphometric estimates, such as regional cortical thickness and brain volumes, are typically estimated from a single scan. Any given morphometric estimate will be a combination of the underlying signal that the experiment is seeking to measure as well as the measurement error (Crocker & Algina, 1986). To the extent that the errors of repeated measurements have low autocorrelation, then measurement error can be reduced by pooling estimates together. Specifically, in the ideal scenario of no autocorrelation, the measurement error will diminish in proportion to the square root of the number of measurements (Angoff, 1953; Brown, 1910; Spearman, 1910). Pooling multiple assessments to drive down measurement error is widely utilized across fields. For example, cognitive tests using response time, neuropsychological tests of cognitive ability, as well as personality and mental health assessments often consist of tens or hundreds of trials or items to drive down measurement error and better estimate an underlying psychological construct (Crocker & Algina, 1986; Kuder & Richardson, 1937). In a particularly relevant example, “measurement-burst” designs, where multiple, closely spaced repeated assessments are made have been utilized in behavioral research to stabilize measurements and maximize sensitivity to detect change within individuals (Nesselroade, 1991; Sliwinski, 2008). Similarly, measures of brain function with fMRI (e.g., brain activation or functional connectivity) increase in reliability when derived from datasets that combine multiple measurements from repeat acquisitions (Birn et al., 2013; Elliott et al., 2018; Laumann et al., 2015). However, to date, pooling has been uncommon in structural MRI studies, mainly because standard T₁-weighted scans are long, causing repeated imaging to be burdensome and costly.

Here, we investigated the potential benefits of acquiring multiple rapid T₁-weighted structural MRI scans and pooling morphometric estimates in a sample of younger and older adults that included individuals diagnosed with mild cognitive impairment and Alzheimer’s disease (AD). A critical feature of our study was to compare the measurement error of morphometrics from a single traditional MPRAGE scan using the ADNI protocol (acquisition time = 5’12”) to pooled morphometric estimates from four CS 1.0 mm scans that were collected in succession (total acquisition time = 4’48”). This allowed for a head-to-head comparison of cluster scanning to standard practices while keeping scan time and participant burden similar. We discovered that pooling estimates from multiple rapid scans can increase the precision of morphometric measurements.

Forty-one paid volunteers participated in our study. All younger adults (19–49 years old; n = 12) were recruited from the community. All older adult participants (55–86 years old) were recruited from the Massachusetts Alzheimer’s Disease Research Center (MADRC) at the Massachusetts General Hospital. Older participants were either cognitively unimpaired (Clinical Dementia Rating, CDR = 0; n = 19) or with mild cognitive impairment or mild dementia (CDR = 0.5 or 1; n = 10). Specific clinical diagnoses were mild multi-domain dementia likely due to AD (probable AD dementia with amyloid and tau CSF or PET biomarker confirmation) (AD; n = 5) or amnestic mild cognitive impairment possibly due to AD based on clinical evaluation with MRI but without molecular biomarker confirmation (n = 5). We chose these groups to explore the viability of rapid scans across individuals with varying levels of age-related atrophy, distinct patterns and degrees of neurodegenerative atrophy, as well as potential movement and compliance challenges typical of patient samples. All participants provided written informed consent in accordance with the guidelines of the Institutional Review Board of Mass General Brigham Healthcare. CDR scores (Hughes et al., 1982; Morris, 1993) were obtained from recent clinical or research visits. Due to excessive head motion and poor data quality detected during quality control, one older participant (CDR = 0) was excluded from all analyses. This resulted in a final sample of 40 analyzed participants (24 females; 60.1 ± 20.2 years; age range: 19–86 years; Table 1).

MRI data were collected using a 3 T Siemens MAGNETOM Prisma^fit MRI scanner (Siemens Healthineers AG; Erlangen, Germany) and the vendor’s 32-channel head coil at the Harvard University Center for Brain Science. The ADNI protocols and scanner were certified with the Standardized Centralized Alzheimer’s & Related Dementias Neuroimaging (SCAN) initiative (https://scan.naccdata.org/). During scanning, participants were given the option to watch video clips (e.g., a nature documentary) or to listen to music. Inflatable cushions were used to provide additional hearing protection and to immobilize the participants’ heads. Every 5–10 minutes, participants were given feedback about motion and reminded to stay still.

The study protocol was designed to compare morphometry from a standard three-dimensional T₁-weighted MPRAGE from the ADNI protocol (Weiner et al., 2017) to cluster scanning using extremely rapid scans. Specifically, we compared the ADNI reference T₁-weighted scan to variants of a research application T₁-weighted rapid scan sequence collected with 6-fold compressed-sensing acceleration (CS) (Mussard et al., 2020). To estimate measurement error, all participants completed two scanning sessions on separate days (i.e., test-retest) within a short period (mean time between scans = 7.7 days ± 5.2 days; 1–25 days). Errors were calculated by comparing morphometric estimates from identical sets of scans that were acquired on two separate days and analyzed independently (Session 1 vs. Session 2).

We investigated 6 different T₁-weighted acquisitions: (1) 1.0 mm isotropic ADNI MPRAGE acquisition (5’12” acquisition; pulse repetition time (TR) = 2300 ms; inversion time (TI) = 900 ms; time to echo (TE) = 2.98 ms; flip angle = 9°; field of view = 256 x 240 x 208 mm; acquisition orientation = sagittal; in-plane GRAPPA acceleration = 2) (Weiner et al., 2017), (2) 1.0 mm isotropic CS scans (1’12” acquisition; TR = 2300 ms; TI = 900 ms; TE = 2.96 ms; flip angle = 9°; field of view = 256 × 192 × 240 mm; acquisition orientation = coronal; CS acceleration = 6x), (3) 0.8 mm isotropic compressed-sensing scans (1’49” acquisition; TR = 2300 ms; TI = 900 ms; TE = 3.10 ms; flip angle = 9°; field of view = 256 × 192 × 230 mm; acquisition orientation = coronal; CS acceleration = 6x), (4) 0.9 mm isotropic CS scans (1’26” acquisition; TR = 2300 ms; TI = 900 ms; TE = 3.03 ms; flip angle = 9°; field of view = 259 × 195 × 231 mm; acquisition orientation = coronal; CS acceleration = 6x), (5) 1.1 mm isotropic CS scans (1’01” acquisition; TR = 2300 ms; TI = 900 ms; TE = 2.92 ms; flip angle = 9°; field of view = 246 × 192 × 247 mm; acquisition orientation = coronal; CS acceleration = 6x), and (6) 1.2 mm isotropic CS scans (0’49” acquisition; TR = 2300 ms; TI = 900 ms; TE = 2.86 ms; flip angle = 9°; field of view = 256 × 192 × 230 mm; acquisition orientation = coronal; CS acceleration = 6x). Turbo Factor/Samples-per-TR could be manipulated independently in the CS sequence but was kept approximately constant at a value close to that used in the ADNI scan to avoid effects from differential T1 weighting during the readout train. A coronal acquisition was employed for the CS scans, in contrast to the sagittal acquisitions in the ADNI scan, as piloting revealed that the sagittal acquisition orientation compounded susceptibility-induced artifacts in the orbitofrontal cortex (Hanford et al., 2021).

The scan protocol consisted of one ADNI scan and 16 CS scans (Fig. 1). In the first half of the scanning session, eight CS scans were collected along with the single ADNI scan. Four of the CS scans were 1.0 mm isotropic scans, and four were isotropic scans of varying resolutions (0.8 mm, 0.9 mm, 1.1 mm, and 1.2 mm). Then, each participant was taken out of the scanner for a brief comfort break. After the break, the participant was repositioned, and the scanner was re-shimmed. An identical set of eight CS scans was repeated, yielding a total of 16 CS scans collected on each day. Across participants, the order of the ADNI and rapid scans was counterbalanced in the first half of scanning to account for potential order effects.

All structural images were processed with FreeSurfer version 6.0.1 using the recon-all processing pipeline (Dale et al., 1999; Fischl et al., 1999), with each scan independently processed. All scans were processed through the default recon-all pipeline with the exception of the 0.8 mm CS scans which were processed with the -hires flag to maintain submillimeter resolution during processing. The results from the automated recon-all pipeline were used without manual interventions or edits. Recon-all included volume- and surfaced-based processing. During volume-based processing intensity normalization, skull stripping (Ségonne et al., 2004), and segmentation of regional brain volumes (Fischl et al., 2002) were conducted. Next, a model of the white-matter surface and the pial surface was generated from each scan using the surface-based processing pipeline (Dale et al., 1999; Fischl et al., 1999). Then, morphometric measures were extracted from the standard recon-all outputs. Specifically, we investigated regional subcortical brain volumes, cortical thickness, and gray-to-white matter signal intensity ratio (GWR) measures estimated for each parcel in the Aseg atlas and the Desikan-Killiany atlas (Desikan et al., 2006; Fischl et al., 2004).

Quality control was conducted by visually inspecting all structural images to note motion artifacts, banding, ringing, and blurring. During visual inspection, minor ringing artifacts were detected in the CS scans that were most evident in the coronal plane in the white matter above the lateral ventricles. Visual inspection of automated labeling and estimated pial and gray/white matter surfaces revealed that these minor ringing artifacts did not visibly affect the estimation process for the CS scans, an impression that was previously tested extensively in quantitative analyses and by gray/white boundaries in key brain regions for each scan type (Elliott et al., 2023). In addition, the results of the recon-all pipeline were checked to confirm that automated processing was completed without error. Critically, no estimates were manually adjusted to allow the automated metrics to provide an unbiased estimate of measurement error.

Measurement error was estimated for 152 separate morphometric estimates. These included 16 subcortical volumes (left and right estimates of the amygdala, accumbens/nucleus accumbens, pallidum/globus pallidus, caudate nucleus, hippocampus, putamen, thalamus, and ventral diencephalon volume from the Aseg atlas; Dale et al., 1999), 68 regional cortical thickness measures, and 68 regional GWR measures (all cortical regions from the Desikan-Killiany atlas; Desikan et al., 2006).

For each scan type and morphometric measure, we estimated the proportion of each measure that is due to measurement error (i.e., percent error). The percent error was estimated for each morphometric measure as the absolute difference between each measure estimated from Session 1 and Session 2 divided by the mean of the two measures. Larger percent errors indicate a greater difference between morphometric estimates from each session and a higher proportion of the measurement that is attributable to measurement error (i.e., lower precision).

To investigate cluster scanning, we first processed each scan independently with FreeSurfer’s recon-all pipeline. Then, for each morphometric estimate, we generated pooled estimates by calculating mean morphometric estimates from the multiple CS scans. We calculated the percent error for the pooled morphometric estimate as the absolute difference between the pooled estimate generated from Session 1 and the same pooled estimate generated from Session 2 divided by the mean total size. For example, to estimate the hippocampal error from four CS scans, we estimated the mean of the hippocampal volume estimates from the first four CS scans from Session 1 and the first four CS scans from Session 2 separately. Then, we quantified the error as the difference between those mean estimates and divided by the mean total size to generate a percent error. We refer to this procedure as pooling and to the outcome of pooling as a pooled estimate throughout this paper.

All vertex-wise analyses were conducted in fsaverage space after each scan’s cortical thickness estimates were resampled from native space to the fsaverage mesh using mri_surf2surf. After resampling to a common space, all cortical thickness maps were smoothed to 20 mm FWHM. Vertex-wise measurement errors were calculated using the same procedure as regional morphometric measures.

The ability of pooling to drive down measurement error will be dependent on the number of scans and the amount of autocorrelation between scans. More autocorrelation means correlated errors and fewer benefits of pooling. Within the structure of this study, we conducted three targeted tests of pooling to investigate alternative methods to minimize autocorrelation and drive down measurement error (see Fig. 1). First, the measurement estimates from cluster scanning were directly compared to those obtained from an ADNI scan. Specifically, the pooled estimates from one, two, and four CS 1.0 mm scans (total acquisition time = 4’48”) were compared to the estimate from the single ADNI scan (acquisition time = 5’12”). As four consecutive CS 1.0 mm scans can be acquired in less time than it takes to collect a single ADNI scan, this comparison allowed us to ask how morphometric estimates obtained from a standard “gold-standard” scan compared to novel rapid, pooled estimates from multiple scans that have lower overall scan burden. Order was fully counterbalanced for this key comparison.

Second, we tested whether a break reduces autocorrelation and improves cluster scanning. Multiple CS 1.0 mm scans were collected both before and after the break (Fig. 1). This allowed us to compare measurement errors between pooled estimates from two CS 1.0 mm scans collected consecutively in the same half of scanning (same head position) with pooled estimates derived from two CS 1.0 mm scans when one was collected before and the second collected after the break (multiple head positions). If a meaningful amount of morphometric error is driven by arbitrary differences in head position and scanner shimming, pooling scans across the break relative to pooling within a sequential block would yield a lower measurement error.

Third, we tested whether variation in scan resolution could reduce autocorrelation. Four CS scans of varying resolutions (0.8 mm, 0.9 mm, 1.1 mm, and 1.2 mm) were collected alongside four CS scans of the conventional resolution (1.0 mm) in the first half of scanning. Two of the non-1.0 mm scans occurred before the 1.0 mm scan block and two after (Fig. 1). This design allowed exploration of the effect of variation in scan resolution on measurement error by comparing pooled estimates from four CS scans of the same resolution with four CS scans of mixed resolutions (i.e., multi-resolution). The average voxel dimension of both groups of CS scans was 1.0 mm and their average ordering in the scan session was equivalent, so order effects were mitigated (but not fully counterbalanced).

Finally, given that eight CS 1.0 mm scans were collected on each day, the design also allowed the pooling of scans from five to eight acquisitions to be explored. This final analysis asked whether it was possible to further drive down measurement error by increasing scan time beyond that of a typical ADNI scan. Eight CS 1.0 mm scans require 9’36” of scan time.

Pooling estimates from multiple 1.0 mm CS scans reduced measurement error. To illustrate error reduction, Figure 2 plots a subset of measures from regions that were chosen because of their importance to brain aging and neurodegenerative disease. For volume and thickness measures, the rate of error reduction was near the rate that would be expected if error variance was unstructured (i.e., proportionate to the square root of the number of scans). For GWR measures, the reductions in error from pooling were somewhat muted, suggesting higher autocorrelation between repeated measurements. Figure 3 comprehensively compares the morphometric measurement error from single and pooled CS scans to their equivalent estimates from ADNI, including every obtained volume, thickness, and GWR morphometric estimate. Errors for all morphometric estimates are reported in the Supplemental Materials.

The mean error for morphometrics from a single CS scan was on average 3.01% (SD = 3.21%) as compared to a slightly smaller mean error for ADNI (M = 2.87%, SD = 3.28%). Already revealing the benefit of pooling, the mean error from two CS scans was 2.32% (SD = 2.36%), which is smaller than the error from ADNI. 77% of the individual morphometric estimates had lower measurement error than ADNI and this difference in error was significantly lower (p < .05) for 20% of the morphometric estimates. Thus, pooling estimates from just two CS 1.0 mm scans resulted in higher precision than ADNI for most morphometric estimates while taking less than half the scan time to collect (2’24” compared to 5’12”).

The mean error for pooled estimates from four CS scans was 1.89% (SD = 1.91%). This reduction in error reflects a 34% lower absolute error than ADNI. At this level of pooling, 93% of individual morphometric estimates had lower measurement error than ADNI, and the difference was significantly lower (p < .05) for 45% of the morphometric estimates. Thus, by pooling estimates from four CS 1.0 mm scans, error was improved for most morphometric estimates while acquisition time was still shorter than the ADNI scan (4’48” compared to 5’12”).

Next, we explored whether the benefits of pooling continued out to eight CS scans. The mean error for pooled estimates from eight CS scans was 1.42% (SD = 1.43%). Such a benefit reflects a 51% lower error than ADNI. 99% of the morphometric estimates had lower measurement error than ADNI and the difference was significantly lower (p < .05) for 86% of the morphometric estimates. However, the rate of improvement was less than would have been expected if the noise in each scan was entirely uncorrelated (51% relative reduction observed vs. 65% expected if error fell at the square root of the number of scans). Thus, pooling reduced error substantially; however, some autocorrelation was likely present, and this dampened the benefit.

Pooling improved the measurement of vertex-wise cortical thickness estimates (Fig. 4). The mean error across all vertices was 1.95% (SD = 1.24%) for ADNI compared to 2.59% (SD = 1.24%) for a single CS scan. As seen in Figure 4, the errors were not uniform and were higher in regions that have low SNR because of their distance to the head coil or due to signal dropout from susceptibility effects (including the cingulate, medial occipital cortex, and the medial and anterior temporal lobes). Paralleling the results from the regional morphometric estimates, the mean error across vertices fell to 2.02% (SD = 0.95%) for two CS scans, 1.59% (SD = 0.70%) for four CS scans, and 1.12% (SD = 0.52%) for eight CS scans. Thus, pooling led to error reductions across the cortex.

To be generally useful, pooling needs to improve morphometric precision in patient populations and in groups where atrophy and poor data quality are particularly challenging. The present results investigated pooling separately for the three subgroups within our study: younger adults, cognitively unimpaired older adults, and older adults with MCI or early-stage AD (MCI/AD; see Fig. 5). All groups showed benefit.

Across morphometric estimates, the mean measurement error for ADNI scans for younger adults was 2.45% (SD = 2.78%), for older adults 2.74% (SD = 2.85%), and for MCI/AD 3.59% (SD = 4.29%). Thus, for the ADNI scan estimates measurement error increased considerably as the sample included more challenging populations. Paralleling the pattern observed for ADNI scans, the mean morphometric measurement error for a single CS scan for younger adults was 2.53% (SD = 2.67%), for older adults 2.91% (SD = 3.13%), and for MCI/AD 3.76% (SD = 3.67%).

The benefits of pooling were independently observed and nearly equivalent across groups. On average, the mean error for pooled estimates from two CS scans for younger adults was 2.04% (SD = 2.01%), for older adults 2.18% (SD = 2.12%), and for MCI/AD 2.91% (SD = 2.97%). Thus, in all groups tested, pooling just two CS scans led to a lower mean error relative to ADNI. The benefit was about a 20% improvement.

The benefit of pooling four CS scans was even greater and also observed independently in each of the groups tested. On average, the mean error for pooled estimates from four CS scans for younger adults was 1.64% (SD = 1.56%), for older adults 1.78% (SD = 1.72%), and for MCI/AD 2.38% (SD = 2.45%). In terms of percentage improvement, this equates to error reductions of 33%, 34%, and 35% for the younger, older, and MCI/AD groups respectively. The benefits of pooling continued to be consistent across groups with eight CS scans. On average, the mean error for pooled estimates from eight CS scans for younger adults was 1.23% (SD = 1.13%), for older adults 1.32% (SD = 1.31%), and for MCI/AD 1.81% (SD = 1.82%). This equates to error reductions of 50%, 52%, and 50% for the younger, older, and MCI/AD groups respectively.

Overall, these subgroup results indicate that despite different levels of absolute error across groups, the benefits of pooling were consistent in all groups and all levels of pooling that we investigated.

Pooling estimates across multiple scans consistently reduced measurement error but at a rate slightly below what would be expected if error variance was unstructured, suggesting that there was autocorrelation between scans contributing to the pooled estimates. Next, we investigated whether a short scan break could reduce this autocorrelation. Halfway through the scan protocol, participants were removed from the scanner and given a short break during which head padding was reset, and the scanner was re-shimmed. If this break reduced autocorrelation, then pooled estimates from scan pairs collected before and after the break should have lower measurement errors compared with two scans collected serially within the same head position. We did not find evidence here for improvement. Across all morphometric estimates, the mean measurement error for two CS scans without a break was 2.32% (SD = 2.27%) and with a break was 2.15% (SD = 2.36%) (Fig. 6). 73% of morphometrics had smaller errors when pooling with a break than without a break. On average, the benefit when pooling two CS scans was a 7% reduction in error and the benefit of a break was significantly lower for only 11% of morphometrics (p < .05). Thus, the benefit of adding break with head repositioning was minimal.

Next, we compared pooled estimates from four CS 1.0 mm CS scans to pooled estimates from four CS scans that combined across multiple resolutions (0.8 mm, 0.9 mm, 1.1 mm, and 1.2 mm isotropic; multi-resolution). These pooled scans were approximately matched in their acquisition time (four CS 1.0 mm scans = 4’48”; four multi-resolution scans = 5’05”). Across all morphometric estimates, the mean measurement error for four CS 1.0 mm scans was 1.89% (SD = 1.91%) and for four multi-resolution CS scans was 1.80% (SD = 1.81%). 69% of morphometrics had smaller errors with multi-resolution pooling. On average, multi-resolution pooling resulted in a 5% reduction in error and the benefit of multi-resolution pooling was significantly lower for 7% of morphometrics (p < .05), reflecting a result near to the false positive rate.

We thus did not find evidence to suggest that designing studies to include a break or multi-resolution scanning would improve precision compared with pooling serially collected 1.0 mm CS scans.

Acquiring multiple fast scans in rapid succession—cluster scanning—achieves more precise morphometric estimates than a standard structural scan. In a test-retest sample of younger and older adults that included individuals with neurodegenerative disease, we found that pooled estimates from four CS scans had substantially less measurement error than estimates from a single traditional scan despite being slightly faster to collect. This was true for regional volume, cortical thickness, and GWR estimates as well as for vertex-wise cortical thickness. The benefits of cluster scanning were clear and consistent in each group investigated: younger adults, older adults, and individuals with MCI/AD. We did not find evidence that pooling CS scans from before and after a break or pooling scans of different resolutions provided further benefit over pooling serially collected 1.0 mm scans. For many purposes, cluster scanning can be considered a viable alternative to the standard 5–8-minute-long structural scan. Cluster scanning yields improved precision and greater flexibility without requiring additional scan time or participant burden.

The first potential benefit of cluster scanning is the improvement of measurement precision (Nielsen et al., 2019). Compared to the ADNI scan, two rapid scans (total acquisition time = 2’24”) had morphometric measurement errors that were 19% lower, while cutting scan time in half (ADNI acquisition time = 5’12”). Studies seeking to limit participant burden or preserve time for additional types of imaging could immediately benefit from this possibility. With four rapid scans, the measurement errors were 34% lower than ADNI despite being slightly faster to collect. This result suggests that studies with a fixed scan time can benefit from the increased precision and flexibility of rapid scans without altering the structure and burden of the study. With eight rapid scans, morphometric precision improved with errors that were half the size of ADNI (51% smaller). This further error reduction came at the cost of additional scan time (total acquisition time = 9’36”). When high precision is critical, such as in longitudinal intervention studies, and additional scan time is tolerable, cluster scanning of eight or more rapid scans might increase sensitivity to detect small changes within individuals.

The most general implication of our findings is that studies implementing cluster scanning might have greater power to detect longitudinal change. Longitudinal studies of brain development, brain aging, and neurodegenerative disease are especially challenging because change is difficult to detect over short periods of time. For example, the annual hippocampal atrophy rate in AD is 3–6% compared to 1–2% in cognitively unimpaired older adults (Barnes et al., 2009; Pini et al., 2016; Schuff et al., 2008) and the amount of measurement error for hippocampal volumes is ~2.5% with the ADNI scan. Therefore, the amount of hippocampal change that occurs in a year is around the size of the measurement error. Large samples and long follow-up times are required to robustly detect group differences and to assess the efficacy of interventions (Aisen et al., 2022; van Dyck et al., 2023; Zetterberg & Bendlin, 2021). Higher precision MRI biomarkers might accelerate research by reducing the sample size and shortening longitudinal study duration. However, longitudinal cluster scanning studies are needed to directly evaluate this possibility. To this point, we have demonstrated that cluster scanning reduces measurement error, but we have not yet confirmed that this increased precision translates to better sensitivity to detect longitudinal change.

Another potential benefit of cluster scanning is that individual scans are quick. Multiple acquisitions provide multiple chances to obtain a usable acquisition. Many of the most important groups participating in basic and clinical brain research include children and individuals with illness that have difficulty with scan adherence and move during MRI scanning. Structural data from these groups are more likely to be affected by poor data quality (Dosenbach et al., 2017; Greene et al., 2018; Reuter et al., 2015). Cluster scanning provides multiple opportunities to address this challenge. First, rapid scans are faster. Each rapid scan is less likely to be corrupted by head motion because there is less time for movement to occur. Second, if a scan is corrupted by head motion, rapid scans can be more easily repeated. Instead of 5–8 minutes of scan time to collect an additional scan, only 1 minute is required. Third, in most scan protocols, when a standard scan is corrupted by motion and the scan cannot be repeated, that participant must be excluded from morphometric analyses because head motion can bias morphometric estimates and lead to spurious inferences (Reuter et al., 2015). This is inefficient and undermines generalizability because participants who move tend to be different from those who are still (Greene et al., 2018; Makowski et al., 2019; Pollak et al., 2023; Reuter et al., 2015). When multiple rapid scans are collected from each participant, morphometric analyses can move forward even when individual scans are unusable because of motion. Fourth, rapid scans can be interspersed throughout a multi-modal scan session so that motion contained to one portion of the scan session will not corrupt all rapid scans. Future studies are needed to explore the benefits of cluster scanning in high-motion groups as compared to alternative approaches such as motion-tracking and dynamic reacquisition (Brown et al., 2010; Maclaren et al., 2013; Tisdall et al., 2012; Zaitsev et al., 2006).

Extreme, rapid scanning is still an emerging area of investigation with multiple options to accelerate scanning and multiple decisions during reconstruction (Bilgic et al., 2014; Mair et al., 2012, 2020; Mussard et al., 2020; Polak et al., 2018). For example, within the rapid scanning option employed in the present comparisons, CS acquisitions require several scan and reconstruction parameters to be manually chosen, including the amount of regularization and the degree of acceleration. We chose these parameters based on extensive piloting (Hanford et al., 2021; Mair et al., 2020; Nielsen et al., 2019); however further optimization is possible, and the optimal parameters may vary for distinct scanner models. Technical advances in reconstruction techniques, including reconstructions with deep learning, promise to provide additional reconstruction options that may further improve in signal-to-noise ratio for images from rapid scanning and the precision of cluster scanning (Hammernik et al., 2018; Knoll et al., 2020). However, for the moment, researchers should pilot and evaluate performance before adopting compressed sensing at a new site and for a new study.

Second, we evaluated cluster scanning in adults across a broad age range and in individuals with MCI and AD. These groups span a range from compliant to challenging participant groups. However, further research is needed to evaluate cluster scanning in additional participant groups, in particular in children and individuals with neuropsychiatric illness.

Third, we investigated many commonly used morphometrics derived from FreeSurfer. However, additional study is needed to test whether cluster scanning provides similar benefits for other morphometric pipelines (e.g., ANTs, FIRST, voxel-based morphometry, and estimation of the boundary-shift-integral; Ashburner, 2009; Freeborough & Fox, 1997; Patenaude et al., 2011; Tustison et al., 2014).

Fourth, we only evaluated cluster scanning with T₁-weighted morphometrics and did not investigate the potential benefits of implementing cluster scanning with other imaging modalities, including T₂-weighted, diffusion, and FLAIR imaging (e.g., Bilgic et al., 2012; Lustig et al., 2007). Future research is needed to investigate cluster scanning and pooling in studies that combine T₁- and T₂-weighted images for morphometric estimation and use other MRI modalities to measure additional metrics (e.g., fractional anisotropy and white matter hyperintensities).

Cluster scanning is a novel strategy for morphometric studies that offers several potential advantages over collecting a single 5–8-minute structural scan. We compared morphometric estimates from the ADNI scan with pooled morphometrics from multiple rapid CS scans. Compared to ADNI, we found that cluster scanning provides the same morphometric precision in less time, and improved precision in the same total amount of scan time. Cluster scanning provides a framework that can adapt to the needs of many studies to maximize scan precision and create additional flexibility in protocol design.

MRI data used in this manuscript will be made available on openneuro.org and the analysis code are available at github.com/maxwellelliott/cluster-pooling.

Maxwell L. Elliott: Conceptualization, Methodology, Software, Investigation, Data curation, Writing—original draft, Visualization, and Funding acquisition. Jared A. Nielsen: Methodology, Software, Writing—review & editing, and Visualization. Lindsay C. Hanford: Methodology, Software, Writing—review & editing, and Visualization. Aya Hamadeh: Resources, Investigation, and Writing—review & editing. Tom Hilbert: Resources, Methodology, Software, and Writing—review & editing. Tobias Kober: Resources, Methodology, Software, and Writing—review& editing. Bradford C. Dickerson: Resources, Writing—review & editing. Bradley T. Hyman: Resources, Writing—review & editing. Ross W. Mair: Resources, Methodology, Data curation, and Writing—review & editing. Mark C. Eldaief: Resources, Writing—review & editing. Randy L. Buckner: Conceptualization, Methodology, Resources, Data curation, Writing—review & editing, Supervision, Project administration, and Funding acquisition.

The study protocol was approved by the Institutional Review Board of Mass General Brigham Healthcare. All participants provided written informed consent in accordance with the guidelines of the Institutional Review Board of Mass General Brigham Healthcare and were compensated.

Tom Hilbert and Tobias Kober are employed by Siemens Healthineers International AG, Switzerland. The authors have no other conflicts of interest to report.

We thank the participants in this research, without whose effort it would not have been possible to accomplish this study. This research was supported by NIA grants R01AG067420, R01DC014296, and P30AG062421, NIH Shared Instrumentation Grant S10OD020039, and Simons Foundation grant 811255. Maxwell L. Elliott is supported by NIA grant K00AG068432. We thank the Harvard Center for Brain Science neuroimaging facility and the Martinos Center Compute Cluster. We thank Kayla Ntoh, Stephanie Kaiser, Joanna Ladoupolou, Francesca Davy-Falconi, and Rachel Lemley for assisting in data collection. We thank Timothy O’Keefe for assisting in data management. We thank Siemens Healthineers for providing the compressed sensing under a research application sequence research agreement. We thank the personnel of the Massachusetts Alzheimer’s Disease Research Center and the Frontotemporal Disorders Unit for their assistance in recruitment, and most importantly the participants and their families for their time and commitment.

Supplementary material for this article is available with the online version here: https://doi.org/10.1162/imag_a_00175