White matter lesions characterise brain involvement in moderate to severe chronic obstructive pulmonary disease, but cerebral atrophy does not

Data were obtained from 31 COPD patients recruited as part of a previous study [5]. Data from six patients were not available at the time of our previous report [5] and have since been included in the dataset. All participants were outpatients recruited from St George’s University Hospital and Royal Brompton Hospital between 2010 and 2011. Seventeen had not been hospitalised within the preceding 12 months. The remaining 14 had previously been inpatients admitted to St. George’s Hospital NHS Trust with a primary diagnosis of COPD exacerbation from whom data were obtained within 12 months of discharge. All participants were assessed whilst in a stable condition. Additionally, 26 controls were recruited from the local community; two of whom were later excluded, one due to a scanner fault and one due to the presence of additional neuropathology (Additional file 1: Table S1, for a complete list of exclusion criteria). Patients were age and gender-matched with controls. They were on average normocapnic at the point of assessment, mildly hypoxaemic and significantly more anxious and depressed than controls. They also had greater pack years smoked, more comorbidities, and lower cognitive function, but did not meet the criteria for dementia (see Table 1).

This study was approved by Wandsworth and East Central London Research Ethics Committees (Ref: 10/H0721/16) and by St George’s University of London, Joint Research Office (Ref: 090147). All subjects gave written informed consent for participation in the study.

Post-bronchodilator spirometry, arterial blood gas analysis, Framingham Stroke Risk Profile (FRSP) [18], Charlson Co-morbidity Index [19], St. George’s Respiratory Questionnaire (SGRQ) [20], and the Hospital Anxiety and Depression scale (HADs) [21] were administered to the patient group. All participants underwent neuropsychological assessment including the Mini Mental State Examination (MMSE), Wechsler Test of Adult Reading (WTAR) (providing an estimate for pre-morbid IQ), and specific sub-scales from the Wechsler Adult Intelligence Scale –III (WAIS-III), Wechsler Memory Scale – III (WMS-III), Delis-Kaplan Executive Function System (D-KEFS), and Rey-Complex Figure Test and Recognition Trial (RCFT) (see [5] for the specific subtests used). Composite scores were calculated assessing the following cognitive domains: Executive Function (average of D-KEFS scaled scores), Episodic Memory (combined average of WMS-III and RCFT scaled scores), Processing Speed (Processing Speed Index from the WAIS-III), and Working Memory (Working Memory Index from the WAIS-III) [5].

Anatomical brain magnetic resonance images were obtained as part of a larger imaging protocol using a 3-Tesla Philips Achieva Dual TX scanner equipped with a 32-channel head coil and gradients up to a maximum of 80 mT/m. Sagittal T1-weighted 3D volume (T1W) images were acquired using a Turbo Field Echo sequence (TE = 3700 ms, TR = 8200 ms, flip angle = 8°, 160 contiguous sagittal slices with an isotropic voxel dimension of 1 mm³ and field-of-view (FOV) of 230 × 182 × 180 mm³). Axial fluid Attenuation Inversion Recovery (FLAIR) images were acquired using a standard inversion recovery sequence (TE = 125 ms, TR = 11000 ms, TI = 2800 ms with 60 contiguous axial slices of 3 mm slice thickness, FOV = 240 × 240 mm² and voxel dimension 0.96²x3mm³).

In view of the need to demonstrate the robust methodology used for this analysis, detailed description of the image analysis procedure is provided.

The conventional Statistical Parametric Mapping (SPM Version 12) [22], geodesic shooting segmentation and normalisation procedure was adapted for use with the present cohort. This pipeline consists of five steps and is described in full in [23, 24]. This technique provides GM, WM, CSF and WML tissue segmentations for each participant.

The voxel-based cortical thickness (VBCT) toolbox in SPM (SPM, Version 12) [22] was used to calculate cortical thickness using the repaired segmentations described above. This procedure is fully described elsewhere [26, 27]. Additionally, the automatic surface-based FreeSurfer pipeline (Freesurfer, Version 5.3.0) [28] was implemented and used to compute vertex-wise pial surface area maps [29] from the T1W images. This was subsequently used to calculate the local gyrification index, defined as the ratio of the pial surface area to that of the perimeter of the brain [30].

In the whole-brain, the GM, WM and CSF segmentations were thresholded at a tissue probability greater than 0.2 and supratentorial cerebral regions were manually extracted. Supratentorial segmentations were summed across all voxels to provide total GM, WM and CSF volumes. Total intracranial volume (TIV) was calculated as the sum of GM, WM and CSF volumes, and the tissue volume ratio was calculated as the sum of GM and WM volume divided by total intracranial volume. Average cortical thickness was computed across the whole-brain. The total number, volume, and average size of WMLs were calculated from the binary WML maps. GM, WM and WMLs volumes were subsequently normalised with respect to head size (calculated as a percentage of total intracranial volume).

The hippocampi were automatically extracted from the T1W images using the standard FreeSurfer ‘recon-all’ pipeline (FreeSurfer, Version 5.3.0) [28, 31, 32]. Segmentation errors were manually corrected using ITK-SNAP (ITK-SNAP, Version 3.2) [33]. Hippocampal volume was calculated for left and right hippocampi separately and normalised with respect to head size as above.

For regional measurements, lobe and deep-GM atlas labels distributed within the ‘Minc Tool Kit’ (Minc Tool Kit, Version 2.2.0) [34] were aligned with the T1W and FLAIR images using Advanced Normalization Tools (ANTs) [35]. The deep-GM atlas label is a composite region comprising the thalamus, caudate nucleus (head and body), putamen and globus pallidus. This was performed by co-registering the FLAIR image to the T1W using a rigid transformation and non-linearly warping each subject’s T1W image to the 1 mm isotropic ICBM 2009c Nonlinear Symmetric T1W standard-space template [36]. The inverse of the transformation was applied to the lobe atlas to bring it into alignment with each subject’s T1W and FLAIR image. WMLs were assigned to the lobe that they maximally overlapped. WM and GM volumes, average cortical thickness, tissue volume ratio and WML, volume, number and average size were calculated for each lobe and deep-GM region.

Whole-brain, lobe and hippocampal statistical analyses were performed using SPSS (IBM Statistical Package for the Social Sciences, Version 24) [37] and FSL’s ’randomise’ (FSL, Version 5.0.6) [38]. Whole-brain and lobe measures were compared between subject groups using ANCOVA models (for Gaussian data and data that could be log₁₀ transformed to Gaussian distributions) or purmutation general linear models (for non-Gaussian data that could not be log₁₀ transformed to a Gaussian distribution). Results from the analysis of lobes were subsequently Bonferroni corrected for multiple comparisons. Between-group differences in hippocampal volume and the interaction with hippocampal hemisphere were tested using a two-way repeated measures design with hemisphere entered as a within-subject effect. Within-group correlations with cognitive and disease severity indices were performed for all whole-brain and hippocampal measures, using partial Spearman’s rank correlations. Correlation results were Bonferroni corrected for the number of statistical tests made for each cognitive or disease severity measure.

All whole-brain, lobe and hippocampal between-group difference and within-group correlative models included age and gender entered as covariates of no interest, hereafter known as confounders. Total intracranial volume was also included as a confounder in any model involving cortical thickness. Estimated premorbid IQ was included as a confounder in any correlative model that tested relationships with cognition. Further within-group correlative models were tested, with pack years smoked and total HADs score entered as additional confounders - in order to determine whether smoking, or depression and anxiety could account for correlation results. Pack years smoked and total HADs score were strongly dependent on subject group and therefore were not suitable to use as confounders in between-group analyses.

To enable voxel-wise statistical analysis the repaired GM segmentations, cortical thickness maps and binary WML maps were warped to the optimised group average template using the deformation fields created previously. The vertex-wise local gyrification index and surface area maps were inflated and registered to the FreeSurfer spherical atlas [32] (FreeSurfer, Version 5.3.0) [28].

Voxel-wise analysis of GM was performed using voxel-based morphometry (VBM) [39]. The standard-space repaired GM segmentations produced previously were modulated by the Jacobian determinant and smoothed using a 6 mm full-width half maximum (FWHM) Gaussian kernel. Smoothed (6 mm FWHM) warped weighted cortical thickness maps were produced using the voxel-based quantification (VBQ) approach described by Draganski et al. [40] and Hutton et al. [27]. Vertex-wise surface area and local gyrification index maps were smoothed using a 6 mm FWHM kernel. WML maps were downsampled by a factor of two in the axial plane to decrease image resolution and increase voxel-wise WML overlap between subjects prior to statistical analysis.

Voxel-wise GM maps were analysed using SPM (SPM, Version 12) [22] and WML maps using the non-parametric mapping (NPM) toolbox distributed within MRIcron (MRIcron, Version 6) [41]. Vertex-wise statistical analyses of cortical thickness and local gyrification index data were performed in FreeSurfer (FreeSurfer, Version 5.3.0) [28]. Group differences and within-group correlations with cognitive and disease severity measures were tested for GM, cortical thickness, surface area and local gyrification index using general linear models. Group differences in WML density were assessed in voxels where WMLs were present in at least 10% of subjects using the liebermeister test [42]. Statistical inference for all voxel-wise and vertex-wise analyses was performed using random-field familywise error (FWE) at p < 0.05. This was implemented using SPM (SPM, Version 12) [22] for GM and cortical thickness analyses, FSL (FSL, Version 5.0.6) for WML and the SurfStat toolbox [43] for surface area and local gyrification.

All voxel-wise and vertex-wise between-group and within-group correlative models included age, gender and total intracranial volume as confounders, except for the voxel-wise WML analysis. Estimated pre-morbid IQ was entered as a confounder in correlative models testing volumetric relationships with cognition. Additional within-group correlative models were tested with pack years smoked and total HADs score included as confounders.

There were no significant differences between patient and control groups in total intracranial volume, tissue volume ratio, whole-brain normalised GM volume, normalised WM volume, average cortical thickness or WML number (Tables 2 and 3). However, COPD patients had significantly greater normalised volume and average size of WMLs than controls (Table 3 and Fig. 2).

Two-way repeated measures ANCOVA of normalised hippocampal volume indicated no significant main effects of hemisphere (F(1,51) = 0.84, p = 0.37) or subject group (F(1,51) = 3.39, p = 0.07), or interaction between group and hemisphere (F(1,51) = 0.11, p = 0.74).

There were no significant differences between patient and control groups for any of the lobe or deep-GM measures (Additional file 2: Table S2).

All voxel-wise and vertex-wise GM measures showed no significant differences between patient and control groups.

No significant differences were found between patient and control groups in WML density although there was a trend for COPD patients to have greater WML density in 88.2% of analysed voxels. The spatial distribution of WMLs was qualitatively similar between patients and controls, with WMLs situated predominantly periventricularly, forming ’caps’ over the anterior and posterior horns, and ’bands’ stretching superior to the body of the lateral ventricles (see Fig. 3 for average WML maps).

There was a significant negative association between patient WML volume and episodic memory, (r _s  =−0.51, p = 0.045) such that patients with greater volumes of WMLs had worse cognitive function (Additional file 2: Table S3). However, inclusion of pack years smoked or total HADs score as confounders in the statistical model removed this association (r _s  =−0.51, p = 0.055, and r _s  =−0.51, p = 0.06, respectively). For patients and controls, there were no further correlations between whole-brain measures and cognitive function that were significant following Bonferroni correction for multiple comparisons (Table 4 and Additional file 2: Table S3).

With respect to COPD severity, the only correlation that survived Bonferroni correction was that between patient total normalised hippocampal volume, and self-reported exacerbation frequency (r _s  =−0.51, p = 0.04) (Table 4). Counterintuitively, it was found that patients who reported having a greater number of exacerbations within the preceding year had larger hippocampi. However, this unlikely result may reflect an inaccuracy of patient self-reporting of exacerbation frequency [44]. The relationship between WML number and FEV₁ % pred. also approached significance such that patients with worse lung function had greater numbers of WMLs (r _s  = 0.50, p = 0.06).

There were no other significant correlations between disease severity indices and any other measures, including the voxel-wise and vertex-wise measures. All findings were unaffected by inclusion of pack years smoked or total HADs score as confounders in the statistical models.

Consistent with the previous literature, we found no evidence of generalised cerebral atrophy occurring in COPD [5, 6, 11, 14]. The lack of voxel-wise and vertex-wise group differences in GM measures is in keeping with the results of Ryu et al. [6] who also reported no localised GM density differences in a small group of COPD patients (19 COPD subjects) with sub-clinical cognitive impairment compared to controls. However, they do not support reports of local GM density reductions in a large COPD cohort (60 COPD subjects) with sub-clinical cognitive impairment [14]. Our findings are also not consistent with evidence for localised GM loss in COPD cohorts with clinical cognitive impairment, for example, in reports of hippocampal atrophy [11, 12], reduction of cortical surface area [13], widespread cortical thinning [13], and localised GM density reductions, predominantly in frontal, limbic and paralimbic structures [7, 8, 11]. This discrepancy in GM findings may relate to cohort differences in severity of cognitive impairment. Any GM reductions in sub-clinical cognitive cohorts are likely to be subtle and consequently the effect sizes may be weaker and group differences less readily detectable.

We found no associations between disease severity and GM measures, however, several studies that reported local GM differences also found associations, particularly that reduced GM volume was correlated with lowered arterial blood oxygenation (indicated by PaO₂ [7, 8, 12] and SaO₂ [12, 13]). Additionally, multiple studies have reported relationships between resting hypoxaemia and lowered neuropsychological performance in COPD e.g. [45‐48]. Most of those studies included patients with moderate-severe hypoxia, so the absence of a significant relationship between GM measures and PaO₂ may be due to our cohort being mildly hypoxaemic.

Our finding of greater volumes and average size of WMLs in patients compared to controls is consistent with previous WML results in COPD [5, 9]. Lobe analyses did not reveal regionally specific differences in WM or WML measures, indicating that greater whole-brain WML volumes are a composite result of small increases in average WML number and size across the lobes. The spatial distribution of WMLs followed a similar pattern in patients and controls, but qualitatively extended further into the WM than controls. The general similarity between WML location in COPD patients and controls suggests a similar aetiology, but with COPD representing a more severe case. WMLs are common within healthy elderly populations with a prevalence of 11–21% in adults aged around 64, increasing to 94% in those aged 82 [49]. However, they can also be indicative of pathological conditions such as cerebral small-vessel disease [50]. Histologically, WMLs represent a heterogenous mixture of diffuse myelin rarefaction with relative sparing of the subcortical U fibres, axonal loss, astrogliosis, spongiosis and widening of perivascular spaces [51]. Their exact pathogenesis remains uncertain, however, it is widely presumed to be ischaemic [50] as they are typically situated along arterial border zones [52], their growth can be predicted by blood flow in surrounding tissue [53] and they are commonly associated with vascular risk factors such as hypertension [54], hyperlipidemia [54], smoking [55], and impaired lung function [54]. Evidence of hypoperfusion [56, 57] and anaerobic glycolysis [58] in the brains of COPD patients, coupled with presence of cerebral microbleeds (another feature of cerebral small-vessel disease) [59] suggest that ischaemic processes are occurring in COPD, potentially secondary to the narrowing or occlusion of the small perforating arterioles at the end of the cerebrovascular tree [50]. The high frequency of concomitant vascular risk-factors in COPD, particularly smoking [55] hypertension [46], and reduced lung function [54] likely contribute to the relatively severe WML-burden in this population.

There are several statistical and methodological concerns, particularly for local GM analyses that potentially limit the reliability of previous neuroimaging findings in COPD. Our study adhered to the recommended minimum standards for statistical analysis in neuroimaging studies [15], so our results are unlikely to represent false-positives. This is in comparison to previous studies that provided statistical inference using the voxel-wise false discovery rate (FDR) (e.g. [6‐8, 13]), performed small-volume analyses without multiple comparisons correction (e.g. [11]) and post-hoc regional analyses (e.g. [7, 8, 11, 14]. Each of these techniques potentially inflate Type-1 error rates due to violations of the assumption of statistical independence. However, by applying rigour in our statistical inference we may have potentially inflated the risk of type-2 error.

Tissue segmentation techniques used in the present study were carefully chosen to provide accurate results. Tissue misclassification can occur when pathology is present, particularly when tissue segmentation is performed from a single MRI modality such as T1W images. This has been demonstrated by Levy-Cooperman et al. [16] in a group of healthy elderly adults such that the presence of WMLs erroneously increased whole-brain GM volumes by up to six percent due to misclassification of WMLs as GM. They also found that this effect was sufficient to disguise group differences in GM between individuals with severe WMLs and patients with Alzheimer’s disease. To avoid these problems we applied the tissue segmentation technique of Lambert et al. [23, 24]. This technique uses multimodal MRI data (i.e. T1W and FLAIR). It provides GM, WM, CSF and WML tissue probability maps and takes advantage of the difference in signal intensity in regions of WM pathology between T1W and FLAIR, to ensure that the WMLs are well defined. A further effect of this technique is that the WML tissue probability maps may be used to repair the GM, WM and CSF tissue probability maps for any tissue misclassification caused by the presence of WMLs [22, 24]. This latter step also increases that accuracy of image coregistration to standard space for voxel-wise analysis of local GM differences. Presence of misclassified regions of tissue can lead to distortions and errors in computed transformations to standard space resulting in misalignment with respect to the template image and error in the results [17] but is overcome by the current method.

The WML segmentation technique used in the present study represents a considerable improvement in objectivity for lesion identification in COPD studies when compared to WML severity visual rating scales [9, 60] and manual segmentation [5], despite requiring a rater to determine WML tissue probability thresholds for each individual. Alternative semi-automatic or fully-automatic techniques are available (see [61] for a review). A recently reported WML segmentation technique is the Brain Intensity AbNormality Classification Algorithm (BIANCA) [62] which has the advantage over our approach in that it that it is fully-automated, however it does require a manually-segmented training set. Our approach does not require a manually segmented training set, instead, a WML prior tissue probability map is automatically generated from T1W and FLAIR images. It remains an open question as to which WML segmentation techniques are the most accurate.

The sample size in the present study is relatively small, although comparable to other studies that have found GM reductions in COPD [7, 8, 11, 13]. As a result some analyses, in particular, the voxel-wise and vertex-wise analyses may be underpowered for detecting small differences and may be vulnerable to outliers. This latter possibility was mitigated by using robust statistical techniques. Additionally, there were substantial differences between groups in terms of pack years smoked and HADs scores, meaning correction for these variables in between group analyses was not possible. The inability to adequately control for differences in smoking history represents a study limitation as smoking is a known risk-factor for developing WMLs [55]. This limitation will be addressed in future studies through direct comparison of COPD patients with smoking controls in our laboratory. Presently, it is unknown how well the semi-automatic method of WML segmentation used in this study, will generalise to use with other patient populations, although it has been successfully applied to a large cerebral small-vessel disease cohort [23, 24]; or to other MRI sequences and scanners.