Age differences in head motion and estimates of cortical morphology

Dataset

Data used in the preparation of this work were obtained from the Cambridge Centre for Ageing and Neuroscience (CamCAN) repository, available at http://www.mrc-cbu.cam.ac.uk/datasets/camcan/ (Shafto et al., 2014; Taylor et al., 2017). The CamCAN dataset includes structural and functional MRI data for a sample of 648 adults across the adult lifespan (aged 18–88; Mean (SD) =54.2(18.5)). All participants were cognitively healthy (MMSE >24) and were free of any neurological or serious psychiatric conditions. See Shafto et al. (2014) for additional details about the sample inclusion and exclusion criteria.

A total of eight participants were excluded from further analyses due to problems with cortical reconstruction or gyrification estimation, yielding a final sample size of 640 adults (326 female, 314 male). Height and weight measurements were available for 559 of the 648 participants (280 female, 279 male), additionally allowing for the calculation of body–mass index (BMI) for this subset of participants (also see Ronan et al., 2016).

Structural measures are derived from a T1-weighted volume acquired using a 3 T Siemens Trio MRI scanner with an MPRAGE sequence. Scan parameters were as follows: TR  = 2,250 ms, TE  = 2.99 ms, flip angle  = 9°, voxel size  = 1 × 1 × 1 mm, GRAPPA  = 2, TI  = 900 ms. Head motion was primarily estimated from two fMRI scans, during rest and a movie-watching task. Both scans lasted for 8 min and 40 s (i.e., 520 s total). For the rest scan, participants were instructed to rest with their eyes closed. For the movie scan, participants watched and listened to condensed version of Alfred Hitchcock’s (1961) “Bang! You’re Dead” (Campbell et al., 2015; Hasson et al., 2008). Note that different scan sequences were used for both of these scans, with volumes collected every 1.970 s or 2.470 s for the rest and movie scans, respectively (see Taylor et al., 2017 for more details); both rest and movie scans had the same voxel size, 3 × 3 × 4.44 mm (32 axial slices, 3.7 mm thick, 0.74 mm gap).

Preprocessing of the structural MRI data

The T1-weighted structural MRIs were processed using FreeSurfer v6.0 (https://surfer.nmr.mgh.harvard.edu/) (Dale, Fischl & Sereno, 1999; Fischl, 2012; Fischl & Dale, 2000). Surface meshes and cortical thickness was estimated using the standard processing pipeline, i.e., recon-all, and no manual edits were made to the surfaces. Gyrification was calculated using FreeSurfer, as described in Schaer et al. (2012).

Fractal dimensionality (FD) is a measure of the complexity of a structure and has previously been shown to decrease in relation to aging for cortical (Madan & Kensinger, 2016; Madan & Kensinger, 2018) and subcortical (Madan & Kensinger, 2017a; Madan, 2018) structures and has been shown to have high test-retest reliability (Madan & Kensinger, 2017b). FD was calculated using the calcFD toolbox (http://cmadan.github.io/calcFD/) (Madan & Kensinger, 2016) using the dilation method and filled structures (denoted as FD_f in prior studies). Briefly, FD measures the effective dimensionality of a structure by counting how many grid ‘boxes’ of a particular size are needed to contain a structure; these counts are then contrasted relative to the box sizes in log-space, yielding a scale-invariant measure of the complexity of a structure. This is mathematically calculated as FD =  − Δlog₂(Count)∕Δlog₂(Size), where Size was set to {1, 2, 4, 8, 16} (i.e., powers of 2, ranging from 0 to 4). To correct for the variability in FD estimates associated with the alignment of the box-grid with the structure, a dilation algorithm was used which instead relies on a 3D-convolution operation (convn in MATLAB) as this approach yields more reliable estimates of FD. This computational issue is described mathematical and demonstrated in simulations in Madan & Kensinger (2016), and empirically shown in Madan & Kensinger (2017b). See Madan & Kensinger (2016) and Madan & Kensinger (2018) for additional background on fractal dimensionality and its application to brain imaging data.

Estimates of head motion

Head motion was estimated using two approaches:

(1) Measured as the frame-wise displacement using the three translational and three rotational realignment parameters. Realignment parameters were included as part of the preprocessed fMRI data (Taylor et al., 2017), in the form of the rp_*.txt output generated by the SPM realignment procedure. Rotational displacements were converted from degrees to millimeters by calculating the displacement on the surface of a sphere with a radius of 50 mm (as in Power et al., 2012). Frame-wise displacement was substantially higher between volumes at the beginning of each scan run, so the first five volumes were excluded. This is the same approach to estimating head motion that is commonly used (e.g., Alexander-Bloch et al., 2016; Engelhardt et al., 2017; Power et al., 2012; Savalia et al., 2017).

(2) Estimated directly from the T1-weighted volume as ‘average edge strength’ (AES) (Aksoy et al., 2012; Zacà et al., in press). This approach measures the intensity of contrast at edges within an image. Higher AES values correspond to less motion, with image blurring yielding decreased tissue contrast. AES was calculated using the toolbox provided by Zacà et al. (in press), on the skull-stripped volumes generated as an intermediate stage of the FreeSurfer processing pipeline. AES is calculated on two-dimensional image planes and was performed on each plane orientation (axial, sagittal, and coronal).

Model comparison approach

Effects of head motion on estimates of cortical morphology (thickness, fractal dimensionality, and gyrification) were assessed using a hierarchical regression procedure using MATLAB. Age was first input, followed by BMI (both with and without age), followed by estimates of head motion from each fMRI scan and the related interaction term with age. In total, eight models were examined, as listed in Table 1. Model fitness was assessed using both R² and ΔBIC.

Bayesian Information Criterion, BIC, is a model fitness index that includes a penalty based on the number of free parameters (Schwarz, 1978). Smaller BIC values correspond to better model fits. By convention, two models are considered equivalent if ΔBIC < 2 (Burnham & Anderson, 2004). As BIC values are based on the relevant dependent variable, ΔBIC values are reported relative to the best-performing model (i.e.,  ΔBIC = 0 for the best model considered).

fMRI-estimated head motion

As shown in Fig. 1, older adults head increased head motion relative to younger adults in both the rest and movie scans (rest: r(638) = .351, p < .001; movie: r(638) = .430, p < .001). Head motion was also greater in the rest scan than during the movie watching (t(639) = 23.35, p < .001, Cohen’s d = 0.99, M_diff = 1.528 mm/min). Nonetheless, head motion was correlated between the fMRI scans [r(638) = .484, p < .001]. While this correlation between scans is expected, particularly since both were collected in the same MRI session, studies have provided evidence that head motion during scanning may be a trait (Engelhardt et al., 2017; Hodgson et al., 2017; Zeng et al., 2014). Moreover, this correlation provides additional evidence that motion during the fMRI scans is consistently larger in some individuals than others, suggesting it similarly affected the structural scans more for some individuals than others and appropriate to include as a predictor for the cortical morphology estimates.

As expected based on prior literature (Beyer et al., 2017; Hodgson et al., 2017), head motion was also correlated with body–mass index (BMI) (rest: r(557) = .456, p < .001; movie: r(557) = .335, p < .001) (Fig. 1). While BMI was also correlated with age (r(557) = .274, p < .001), BMI-effects on head motion persisted after accounting for age differences (rest: r_p(555) = .340, p < .001; movie: r_p(555) = .249, p < .001).

While head motion was substantially lower in the movie condition than during rest, it was relatively stable over time (e.g., it does not tend to decrease over time). However, in the movie watching task, there is evidence of systematic stimuli-evoked increases in head motion (Fig. 2), e.g., around 280 s and 360 s. These periods of increased head motion correspond to events within the movie; in the first period, the boy is loading the real gun with bullets, the second, more prominent period is a suspenseful scene where it appears that the boy may accidentally shoot someone. Moreover, these events also correspond to fMRI differences in attentional control and inter-subject synchrony (see Campbell et al., 2015).

T1-estimated head motion

Head motion was also estimated directly from the T1-weighted volume as the average edge strength (AES), following from Zacà et al. (in press); higher AES values correspond to less motion. Here I calculated AES for each plane orientation. AES in the axial and sagittal planes was moderately related to age (axial: r(639) = .493, p < .001; sagittal: r(639) = .525, p < .001) (Fig. 3); AES in the coronal was only weakly correlated with age (r(639) =  − .131, p < .001). AES in the axial and sagittal planes were strongly correlated with each other (r(639) = .702, p < .001).

Interestingly, AES was relatively not related to BMI (all |r|’s <.2). AES was also relatively unrelated to fMRI-estimated head motion (rest: r(639) = .112, p = .005; movie: r(639) = .148, p < .001). Thus, while AES is sensitive to an MR image property related to age, it seems to be distinct from fMRI-estimated head motion. A likely possibility is that AES here is detecting age-related differences in gray/white matter contrast ratio (GWR), as have been previously observed (Knight et al., 2016; Magnaldi et al., 1993; Salat et al., 2009). In contrast, the mechanism for the correlation between BMI and fMRI-estimated head motion is likely apparent–rather than real–head motion caused by respiratory chest motion producing susceptibility variations in the B0 field (Raj, Anderson & Gore, 2001; Van de Moortele et al., 2002; Van Gelderen et al., 2007).

Cortical morphology

As shown in Fig. 4, mean cortical thickness significantly decreased with age (r(638) =  − .652, p < .001, −0.0432 mm/decade), as did fractal dimensionality (r(638) =  − .705, p < .001, −0.0097 FD_f/decade) and gyrification (r(638) =  − .427, p < .001, −0.0372 GI/decade). All three slopes (change in metric per decade) are nearly identical to those first calculated by Madan & Kensinger (2016), as is the finding of higher age-related differences in fractal dimensionality and weaker differences in gyrification (also see Madan & Kensinger, 2018). However, it is also worth acknowledging that AES in the axial and sagittal planes were comparably correlated with age as gyrification. Effects of BMI on all three measures of cortical morphology were relatively weak (thickness: r(557) =  − .169, p < .001; fractal dimensionality: r(557) =  − .168, p < .001; gyrification r(557) =  − .083, p = .049).

Of particular interest, I examined the influence of head motion on the cortical morphology estimates. For all three measures, head motion explained only a small amount of additional variance beyond age, as shown in Table 1. Nonetheless, head motion from the movie scan did explain significant additional variance, as measured by ΔBIC, however, this only accounted for an additional 1% variance in the cortical morphology measures. In the model of cortical thickness including head motion from the movie scan (but not the interaction), age related changes corresponded to −0.0398 mm/decade, while head motion contributed −0.0135 mm/(mm/min).