Increased white matter glycolysis in humans with cerebral small vessel disease

Abstract

White matter lesions in cerebral small vessel disease are related to ischemic injury and increase the risk of stroke and cognitive decline. Pathological changes due to cerebral small vessel disease are increasingly recognized outside of discrete lesions but the metabolic alterations in non-lesional tissue has not been described. Aerobic glycolysis is critical to white matter myelin homeostasis and repair. In this study, we examined cerebral metabolism of glucose and oxygen as well as blood flow in individuals with and without cerebral small vessel disease using multi-tracer positron emission tomography. We show that glycolysis is elevated in non-lesional white matter in individuals with small vessel disease relative to healthy, age-matched controls. On the other hand, in young healthy individuals, glycolysis is relatively low in areas of white matter susceptible to lesion formation. These results suggest that increased white matter glycolysis is a marker of pathology associated with small vessel disease.

Introduction

White matter (WM) lesions (WMLs) owing to cerebral small vessel disease (SVD)¹ manifest as T2-weighted hyperintensities in brain magnetic resonance images (MRI) and are common in the aged population². WMLs are associated with increased risk of stroke, depression, and cognitive impairment^3–5. Mounting pathologic evidence demonstrates endothelial dysfunction and ischemia within senescent WMLs^1,6,7. Thus, studies assessing the structure^8,9 and physiology (e.g., measures of blood flow)^10,11) within WMLs predictably identify correlates of ischemia. WM surrounding WMLs demonstrates abnormal microstructural features^12–14 and subtle T2 signal increases^8,15 suggesting that non-lesional WM might be metabolically abnormal. Normal WM is glycolytic^16,17; oligodendrocyte-derived glycolytic intermediates support axonal metabolism and increased glycolysis supports repair of injured tissue¹⁸. These observations suggest that WM in humans may similarly rely upon glycolysis for metabolic support and protection against injury. This study aims to understand changes in cerebral metabolism in and around WMLs and to understand how such changes relate to WM vulnerability.

Results

Cohort and Imaging Procedures

To investigate the role of glycolysis in human WM in the context of small vessel disease (SVD)¹, we selected data from the AMBR dataset¹⁹, which includes multi-tracer positron emission tomography (PET) and high-resolution structural MRI in humans without and with SVD. Participants were assigned to three groups: healthy, young controls (HC-Y, mean age = 38y, N = 20), healthy, older controls (HC-O, mean age = 64y, N = 18), and participants with WMLs on MRI (mean age = 76y, N = 22). Demographic data are reported in Supplemental Table 1. The HC-Y and HC-O participants were free of radiographic evidence of SVD, allowing for mild periventricular capping or scant, nonspecific punctate WM foci (Fazekas score ≤ 1²⁰). The SVD group had clear WMLs on MRI (Fazekas score > 1). Each participant underwent PET imaging with ¹⁸F-FDG, ¹⁵O-H₂O, ¹⁵O-O₂, and ¹⁵O-CO. We calculated the cerebral metabolic rate of glucose (CMRglc) and oxygen (CMRO₂), cerebral blood flow (CBF), and the glycolytic index (GI), which reflects aerobic glycolysis²¹. The MRI and PET data were spatially co-registered²² and WMLs were segmented on MRI²³. FreeSurfer²⁴ segmented the structural images into gray vs. white matter. Additional methodological detail is given in the Supplemental Online Methods.

Increased Glycolysis in Normal Appearing White Matter of Patients with Cerebral Small Vessel Disease

Partial volume corrected (PVC) metabolic parameters (CMRglc, CMRO₂, CBF, and GI) in different tissue classes (cerebrospinal fluid [CSF], GM, WM, and lesion) were estimated across groups (Figure 1A) using region of interest (ROI) based PVC²⁵. Supratentorial GM was included as a single ROI in this analysis. Standardized uptake value ratios (SUVR) were referenced against the whole cerebellum for each metabolic parameter, tissue class, and participant. Linear models, with participants as repeated measures, applied to the metabolic parameters revealed a significant interaction between group and tissue type (Supplemental Tables 2 and 3). CMRglc (Figure 1B), CBF (Figure 1C), and CMRO₂ (Figure 1D) were greater in GM as compared to WM and severely reduced in WMLs. GM glucose (Figure 1B) and oxygen metabolism (Figure 1D) were lower in the SVD group as compared to both HC-Y and HC-O. Importantly, WM CMRglc was elevated in the SVD group compared to HC-Y and HC-O. CMRglc in lesions (defined only in the SVD group) was near zero indicating very low glucose utilization. WML CBF and CMRO₂ was also substantially lower compared to WM and GM. These results demonstrate that metabolism is severely compromised in WMLs and that SVD alters metabolism in non-lesional WM. Direct measurement of GI in WMLs yielded highly variable results due to the relatively small volume of WMLs and that GI is particularly sensitive to measurement error in the denominator (CMRO₂), which is very low in WMLs. Additional considerations (see Supplemental Figure 1), indicate that changes in GI are not explained by changes in total glucose metabolism.

Figure 1.

Metabolic parameters measured in WM, GM, and Lesion in HC-Y, HC-O, and SVD groups. Panel A shows a schematic overview of tissue segmentation and metabolic parameter estimation. T1-weighted and FLAIR images were spatially aligned. Tissue classes were assigned and sGTM was used to calculate metabolic parameters for each tissue class, relative to the whole cerebellum (SUVR). Lesion metabolism is defined only in the SVD group. Panels B-E show median, Q1, and Q3 values for CMRglc (B), CBF (C), CMRO₂ (D), and GI (E) across tissue classes (WM, GM, Lesion) and groups (HC-Y, HC-O, SVD). SUVR nominally is a positive quantity; negative values arise from imprecision in the matrix inversion step of sGTM and represent very small values. GI is expressed in arbitrary units. Higher or lower values indicate more or less glycolysis in one region compared to another, and zero indicates the whole brain average. Error bars are standard deviation of the mean for each group with individual participants as samples (N HC-Y = 20, N HC-O = 18, N SVD = 22).

As illustrated in Figure 1, WM GI was relatively depressed in HC-O group as compared to the HC-Y group but higher in the SVD group as compared to HC-O. This last finding most likely represents a metabolic response to pathology as opposed to reversal of the effects of aging. SVD is associated with age-related comorbidities such as Alzheimer disease (AD). Since the AMBR dataset was designed to investigate both aging and AD, a minority of SVD participants exhibited evidence of abnormal amyloid-beta accumulation (N=9) or were subsequently found to have mild cognitive impairment defined as Clinical Dementia Rating^® score of 0.5 (N=6), including 5 with evidence for both amyloid accumulation and cognitive impairment. The results in Figure 1 are reproduced in Supplemental Figure 2 with these participants identified. We find no evidence that that this subset of participants drove the observed effects.

Glycolysis is Increased in Peri-lesional Tissue

To investigate how metabolism changes over the peri-lesional zone while also accounting for partial volume effects, we systematically dilated the boundary between lesion and WM (Figure 2A). Our segmentation model computes the probability that each voxel belongs to a different tissue class (CSF, WM, GM, lesion)²³. By adjusting the threshold at which voxels are classified as lesion, the lesion ROI can be made to systematically include increasing proportions of surrounding tissue. High (conservative) lesion thresholds correspond to distinct, discrete lesion voxels whereas low (liberal) thresholds correspond to ROIs including peri-lesional voxels. Metabolic measurements made in these ROIs are then subjected to region-based PVC and presented as a function of lesion threshold (Figure 2B). For CMRglc, CBF, and CMRO₂, lesion metabolic estimates were very low at high lesion thresholds, reflecting the expected properties of distinct WMLs. As the lesion threshold decreases, metabolic characteristics of the WML ROI increasingly resemble WM. The size of this effect, depicted as the slope of the WML estimate in Figure 2B, is greatest for CMRglc and less for CBF and CMRO₂. Hence, GI increases sharply as a function of decreasing lesion threshold.

Figure 2.

Perilesional metabolism resembles normal appearing WM metabolism in SVD. Panel A shows a schematic of the analysis strategy. ROIs were generated by systematically varying the threshold required for a voxel to be assigned to the WML ROI. Lower thresholds correspond to larger WML ROIs with more surrounding non-lesional WM. Metabolic parameters were estimated by sGTM parametric in lesion threshold. Plots derived from that analysis are shown in panel B; the line indicates the mean measurement and error bars indicate one standard deviation across individual subjects as samples (N SVD = 22). As expected, parameter estimates in GM expectedly did not depend on lesion threshold. The derived quantities were entered into a linear model to estimate the contributions of lesion core vs. peri-lesional tissue. Compartmental estimates derived from this model are shown as box plots in panel C. Each sample in the box plot corresponds to an individual SVD patients (N = 22). Center of boxplot corresponds to the median, the lower and upper hinges represents the first and third quartiles, the whiskers represent 1.5 x (Inter-quartile range), and values outside that range are plotted individually.

At a given lesion threshold , metabolic parameters (Figure 2B) reflect the lesion core as well as perilesional tissue. We define the lesion core as voxels with a lesion classification probability >90% (marked hyperintensity on T2-weighted images). Perilesional tissue is defined as voxels with tissue intensities between WMLs and WM. Estimating metabolic parameters directly in perilesional tissue is precluded by 1) lack of a priori boundaries and 2) limited spatial resolution of PET. We used linear modeling to solve for core and perilesional metabolism parametric in lesion threshold. Solving this linear model yielded parameter estimates for core and perilesional tissue (Figure 2C). CMRglc, CBF, and CMRO₂ were lower in lesion cores compared to all other tissue classes, in accordance with the directly estimated values (Figure 1). Perilesional CMRglc, CMRO₂ and CBF were similar to values in normal appearing WM. GI in the peri-lesional zone was also similar or slightly higher when compared to normal appearing WM, which is elevated compared to WM in age-matched healthy controls. Statistical tests of these comparisons are shown in Supplemental Table 4. These results suggest that WMLs are metabolically distinct from the surrounding WM, which exhibits increased glycolysis.

The above-discussed results suggest that extra-lesional WM is glycolytic in SVD. We next sought to validate this result using an alternative approach, specifically, metabolic parameter estimation using Region Based Voxelwise (RBV) partial volume correction (PVC)²⁶, which does not rely on the assumption of a discrete perilesional compartment (Figure 3A). Metabolic parameters were modeled parametric in distance from the lesion-WM boundary (parameter d; Figure 3B). This analysis again showed that CMRglc is very low within lesions but rises sharply outside the lesion to values slightly above the mean WM value, indicated as a dashed blue line. A maximum in estimated CMRglc occurred around 5-10 mm from the lesion boundary. CBF and CMRO₂ behaved in a broadly similar manner. Voxelwise GI was not modelled as it is an unstable ratio owing to noisy denominators. Nonetheless, the results presented in Figure 3B suggest a rapid transition from marked hypometabolism of glucose and oxygen within the lesion to a metabolic profile similar to normal appearing WM within approximately 5 mm of the lesion boundary. The slight overshoot in CMRglc may represent elevated glycolysis in the absence of similar changes in CBF and CMRO₂.

Figure 3.

Spatial topography of WM metabolic parameters. Panel A shows a schematic of the data analysis. Each WM voxel was assigned a distance from lesion boundary (green arrow) or CSF boundary (purple arrow). RBV corrected metabolic parameters then were modeled parametric in distance from lesion edge (panel B). Green lines show metabolic parameter means across individual subjects (thin black lines). The dashed blue line is the estimated metabolic parameter value of normal appearing WM. The same analysis was repeated parametric in distance from the ventricle (magenta line in panel C); results parametric in distance from the lesion boundary are superimposed. Panels A-C are specific to the SVD group. Panel D shows metabolic parameters parametric in distance from CSF boundary in all three groups. Error bars in panels C and D are SEM with individual subjects in each group serving as the sample (N HC-Y = 20, N HC-O = 18, N SVD = 22).

WMLs occur in a stereotypical, predominantly periventricular distribution. It follows that parameter d (distance from lesion boundary) in the preceding analysis is conflated with proximity to the ventricular wall, notwithstanding that CSF voxels were excluded. Prior reports suggest that cerebral blood flow varies depending on distance from the ventricular wall²⁷. To quantify the extent to which this feature is reflected in our data, we repeated the above analysis using distance from ventricular wall instead of distance from lesion edge (Figure 3C). Broadly similar results were obtained when metabolic parameters were estimated parametric in distance from ventricular wall. This similarity is consistent with the limited resolution of PET, as the topography of metabolic changes parametric in distance from lesion vs. ventricle cannot be easily distinguished.

The crucial question remains the extent to which the features shown in Figure 3B–C (decreasing CMRglc, CBF, CMRO₂ near lesions or ventricle) represent normal physiology or sequelae of a pathological process. To address this, we calculated metabolic parameters parametric in distance from the ventricular wall in HC-Y and HC-O, exactly as in the SVD group (Figure 3D). Distance from lesion boundaries is undefined in the control groups as they lack lesions. Whereas CBF and CMRO₂ behaved similarly with respect to distance from ventricle across all groups, CMRglc remained increased in the SVD group relative to HC-Y and HC-O.

Low Levels of Glycolysis Predispose to White Matter Lesion Formation

WMLs tend to occur in a stereotypical distribution surrounding the lateral ventricles, particularly in regions commonly referred to as “caps”^28,29. This phenomenon suggests the possibility that such regions are metabolically distinct from regions of WM less prone to develop lesions. To address this question, we derived a lesion probability map in the SVD group (Figure 4A) and measured the corresponding physiological parameters in HC-Y and HC-O. Specifically, WM in the HC groups was partitioned into two classes: vulnerable WM (WM_v) and non-vulnerable WM (WM_n), based on lesion frequency in SVD. Since lesion frequency is a continuous measure, we performed this analysis over probabilities in the range of 5-30%. In both HC-Y (Figure 4B) and HC-O (Figure 4C), across all lesion frequency thresholds, mean CMRglc, CBF, and CMRO₂ were consistently lower in vulnerable WM as compared to non-vulnerable WM. Importantly, this difference was not significantly modulated by age (comparing Figure 4B vs. Figure 4C). Moreover, CMRglc exhibited the steepest dependence on lesion frequency threshold in comparison to CBF and CMRO₂, indicating that non-vulnerable WM is more glycolytic as compared to vulnerable WM (see also GI, Fig. 4B and 4C). Thus, periventricular WM is metabolically distinct from superficial WM and this difference is present in young and older healthy participants.

Figure 4.

Regions prone to lesion formation are less glycolytic in young and old control participants. Panel A shows the lesion density heat map in the SVD group. The remaining panels show metabolic estimates parametric in the lesion density boundary separating vulnerable (WM_v) vs. non-vulnerable (WM_n) WM. Color shows fraction of patients with lesions at a given voxel. B: HC-Y group. C: HC-O group. The x-axis corresponds to the boundary between WM_n vs. WM_v in units of lesion density. Lines in both B and C indicate the mean; error bars represent +/− one standard deviation with individual subjects in each group serving as samples (N HC-Y = 20, N HC-O = 18).

Discussion

We used PET to investigate glycolysis changes in non-lesional WM and GM associated with SVD. We also assessed glycolysis as a metabolic predisposition to WML formation. Technical advancements in this work include application of multiple, complementary PVC techniques and measurement of aerobic glycolysis in anatomical relation to WMLs. We demonstrate that non-lesional WM is relatively more glycolytic in SVD, that glycolysis is maximally elevated near the lesion boundary, and that lesions tend to form in regions that are weakly glycolytic in healthy controls.

Glycolysis in the brain declines in some individuals with age³⁰. Prior results as well as this study show that WM normally is glycolytic^31,32 but becomes less so with healthy aging. Here, we show that WM glycolysis is relatively elevated in the setting of SVD, suggesting that WM glycolysis represents a protective and/or compensatory mechanism. These findings are concordant with studies in mice showing that glycolysis in oligodendrocytes provides trophic support to axons via glycolysis³³, especially in respiration deficient tissue³⁴. Thus, glycolytic WM surrounding WMLs in humans with SVD is pathophysiologically rational and has a precedent in mouse models.

The mechanisms underlying increased WM glycolysis in patients with SVD remain unclear. It is possible that this tissue has subtly impaired blood flow and oxygenation³⁵, not detectable in our study, leading to increased dependence on glycolysis for generating ATP. Another possibility is that WM in SVD develops glycolysis in response to factors released by adjacent WMLs, such as hypoxia-inducible factors³⁶. Alternatively, WM glycolysis may reflect inflammation or reparative processes, since glycolysis is a feature of activated microglia and other inflammatory cells³⁷. Both neurons and oligodendrocytes react to acute metabolic stress by increasing glycolysis and releasing lactate^38,39; chronic metabolic stress in the setting of SVD may produce a similar adaptive response. Finally, glycolysis may support (re)myelination and axonal repair⁴⁰. These explanations are not mutually exclusive, and more studies are warranted to adjudicate which are correct.

The stereotypical topography of WMLs in SVD has been attributed to decreased cerebral blood flow⁴¹, and sensitivity to hypertensive changes⁴². The present work demonstrates that SVD increases glycolysis relatively in non-lesional WM generally but that specific regions with relative deficiencies of glycolytic metabolism are predisposed to WMLs. This vulnerability could result from decreased protection provided by glycolytic support cells^33,40 or lower repair potential of less glycolytic tissues ¹⁶. Regardless of the mechanism, we hypothesize that glycolysis is protective in the WM of patients with SVD⁴³. If true, interventions to improve white matter glycolysis might prevent SVD progression.

Both MRI and histopathological studies suggest that periventricular and deep WMLs are pathologically distinct⁴⁴. Periventricular lesions are associated with fewer pathological changes compared to large or irregular WMLs within the deep WM, and confluent lesions show more parenchymal destruction including loss of myelin, axons, and oligodendrocytes⁴⁵. The mechanistic basis for this topographic variability is unclear. Prior work has hypothesized that hypoperfusion⁴¹ resulting in hypoxia/ischemia, changes in blood-brain barrier function, or microglial activity, may mediate differences in pathological damage^45,46.

This study has important limitations. First, our analyses are based on relative metabolic measures (SUVR). Thus, metabolic differences in the reference region (cerebellum) could bias the results. The pons has been used as a reference region^47,48 but we did not do so as it is frequently affected by WMLs⁴⁹. Quantitative measurements will be needed to determine whether and how SVD relates to absolute changes in metabolism. Second, we did not identify an age effect in normal-appearing WM CBF whereas prior studies have demonstrated such an effect. This could relate to the separation of older adults with and without SVD or, again, from the use of relative, reference-region based measures. Third, the limited spatial resolution of PET mandates that PVC methods with their attendant assumptions must be used and that the smallest lesions could not be resolved. The relations between WM and GM glycolysis remain unclear^30,32. Higher resolution PET will resolve these questions by accounting for topographies within both WM and GM. Finally, a few among our older adult cohort were found to have evidence of brain amyloid accumulation, mild cognitive impairment, or both. Although the low numbers of such participants preclude formal statistical treatment, the analysis in Supplemental Figure 2 suggests that these participants did not measurably bias the presented results. Nevertheless, there remains the caveat of other potential confounding comorbidities.

In conclusion, we report the existence of WM glycolysis in patients with SVD adjacent to white matter lesions. Moreover, areas less prone to WMLs are characterized by relatively higher levels of glycolysis in healthy, young individuals. Considering the current evidence including that presented here, we hypothesize that, in humans, WM glycolysis is protective against WMLs and might increase in response to injury in a compensatory or reparative role.

Methods

Participants and Regulatory Approvals

All reported results conform with the requirements of the Institutional Review Board and Radioactive Drug Research Committee requirements at Washington University in St. Louis. Each participant provided written informed consent. These data were obtained from the Adult Metabolism & Brain Resilience (AMBR) dataset—a large PET and MRI study of brain aging and Alzheimer’s disease beginning in 2013 and currently ongoing¹⁹. All individuals were recruited from the Washington University in St. Louis community and/or the Knight Alzheimer Disease Research Center. All participants were free of known confounding neurological, psychiatric, or major medical comorbidities that might compromise participation (e.g., poorly controlled diabetes). Nine participants demonstrated increased burden of amyloid-beta based on PET study and six participants were cognitively impaired (CDR > 0) when evaluated in a co-enrolled study. Individuals were excluded if they had contraindications to MRI or PET including possible pregnancy.

Participants were divided into three groups for the purpose of this study: healthy control younger (HC-Y), healthy control older (HC-O) and patients with evidence of cerebral small vessel disease (SVD). Healthy control participants, in addition to having no known major comorbidities, were required to be free of radiographic evidence of SVD allowing for minimal periventricular capping and/or a few nonspecific punctate WM foci (Fazekas score <= 1) ²⁰. Healthy control participants were split into younger (age 25 to 49 years) or older (age 56 to 75 years) groups. SVD patients were required to have radiographic evidence of SVD (Fazekas score > 1)

MRI Acquisition

MRI scans were obtained in all subjects for the purpose of tissue and lesion segmentation. Images were acquired on a Siemens Trio or (retrofitted) Prisma 3 T scanner. A high resolution T1-weighted magnetization-prepared rapid-gradient-echo (MPRAGE) sequence and T2-weigthed fluid attenuated inversion recovery (FLAIR) sequence were acquired in each subject.

PET Acquisition

All participants underwent PET imaging on a Siemens ECAT HR+. All individuals underwent a single ¹⁸F-FDG and two sets of ¹⁵O PET scans. FDG scans were performed after a slow intravenous injection of 5 mCi of FDG. Venous samples of plasma glucose before and midway through the scan verified that glucose levels remained nominal throughout the study. Dynamic acquisition of the images was performed and the last 20 minutes (40 min – 60 min post-injection) were summed to create the CMRglc image.

Each ¹⁵O PET session consisted of two sessions of three scans to determine cerebral blood volume (CBV), cerebral blood flow (CBF), and cerebral metabolic rate of O₂ (CMRO₂). Cerebral blood volume was measured with a 5-minute emission scan beginning 2 minutes after a brief inhalation of 40 to 110 mCi of ¹⁵O-CO. Dynamic scans of 3 minutes duration were acquired following injection of 50 mCi of ¹⁵O-H2O or inhalation of 40 to 110 mCi of ¹⁵O-O₂ in room air for measurement of CBF and CMRO₂, respectively. Whole-brain time-activity curves were constructed and the optimal 40 second frame was summed to create the parametric image. Given their 2-minute half-life, at least 12 minutes elapses between each ¹⁵O injection and prior to the final ¹⁸F-FDG injection. CMRO₂ calculation was calculated as previously described³⁰. Glycolytic index was calculated as the difference between observed CMRglc and CMRglc predicted on the basis of CMRO₂ as previously described²¹. No specific instructions were given to participants other than to remain awake with their eyes closed.

Interpretation of Glycolytic Index (GI)

The GI indexes the local ratio of glucose to oxygen utilization. In quantitative PET, the CMRO₂:CMRglc ratio (the OGI) is 6 if all glucose is oxidatively metabolized to CO2 (to generate ATP) and less than 6 if some fraction of glucose is metabolized via other pathways (e.g., the Warburg effect or the pentose-phosphate shunt) for energetic or non-energetic purposes. Linear regression over the whole brain of voxelwise CMRglc on CMRO2 generates a best fitting line and reveals regions of the brain falling above as well as below the line of regression. Areas above the line of regression are said to be relatively more glycolytic. In non-quantitative PET, both CMRglc and CMRO₂ are evaluated by SUVR, i.e., the ratio of local tracer uptake to a reference region (the cerebellum or whole brain). Accordingly, whereas the location and scale of GI are unknown relative to a true quantitative measure of aerobic glycolysis, there is very high correlation (r>0.9) between GI and other measures of aerobic glycolysis (including OGI²¹), allowing GI to be an accurate yet accessible relative measure of regional aerobic glycolysis. GI is expressed on an arbitrary scale in arbitrary units. Thus, 0 has no intuitive meaning and it follows that positive and negative values also do not have an intuitive meaning, except in a relative sense. The only inference is that there is more or less GI between the two regions.

Image Registration and Tissue Segmentation

Summed ¹⁵O images were mutually aligned and then these images and the ¹⁸F-FDG summed image were aligned to the MPRAGE image. Similarly, the FLAIR image was aligned to the MPRAGE. These alignments were computed using in-house software (4dfp suite of tools) and performed in a single resampling step.

Tissue segmentation was performed using the FreeSurfer²⁴ software into well-defined gray and white matter regions of interested on the basis of the MPRAGE scan. FreeSurfer does not reliably identify T2 hyperintensities (i.e., WMLs) or peri-lesional abnormalities in WM. Thus, we adopted a complementary strategy to identify WMLs and subtle surrounding WM abnormalities²³. MPRAGE and FLAIR images were intensity normalized using a previously described histogram normalization approach. Different tissue classes are defined by distinct loci (MPRAGE, FLAIR) in this intensity space. Thus, for any voxel, the tissue score can be computed which indexes the probability of that voxel belonging to a defined tissue class (CSF, GM, WM, WML). For any one voxel, these values sum to 1.

Metabolic Parameter Estimation in Fixed Regions of Interest (Figure 1)

Estimation of activity attributable to a specific tissue requires partial volume correction owing to the limited spatial resolution of PET. We used a region of interest (ROI) based approach which assumes that each ROI has a different activity level and that this activity level is constant throughout the ROI. The specific algorithm was symmetric geometric transfer matrix (sGTM)²⁵. This algorithm takes a PET image and an ROI image as arguments are returns a scalar estimate of the activity in each ROI. The input PET images were the CMRglc, CBF, and CMRO₂. GI images are very noisy and thus GI was calculated for each ROI following partial volume correction of CMRglc and CMRO₂. ROI images were constructed form the FreeSurfer segmentations and large ROIs for WM and GM were defined for all groups. For the SVD group, an additional WML class was defined as clusters of 30 or more voxels with a WML tissue score > 0.7. This yielded lesion segmentations that capture all large lesions and excluded punctate lesions well below the resolution of PET. Application of sGTM to the PET and ROI images yielded a WM and GM (and WML, in the SVD group) value for each participant. These metabolic parameters were subjected to linear models with group (HC-Y, HC-O, and SVD) and tissue (WM, GM, WML) as factors. Participant ID was included as a random effect. The numerical results of this analysis are shown in Figure 1 and the statistical results are shown in Supplemental Table 2.

Core Lesion and Peri-lesional Metabolism Revealed by Iterative Dilation of Lesion:NAWM Boundaries (Figure 2)

The preceding analysis is predicated on the assumption that there is a discrete boundary between WMLs and NAWM. The following analysis was designed to test the validity of that assumption. Within the SVD group, GM was defined based on the FreeSurfer segmentation. WM was defined by the FreeSurfer segmentation and then regions with WML tissue scores exceeding a threshold were defined as WMLs. This threshold was systematically varied from 0.8 to 0.2. Threshold values above and below this range led numerical instability. To reduce uncertainty in the matrix inversion step of sGTM, voxels defined as WMLs but in connected clusters of less than 20 voxels were defined as WM. For each ROI image resulting from a different WML threshold, sGTM was performed and the resulting metabolic parameters are shown in Figure 2B. This shows that as WML ROIs increase in size, and include more and more NAWM, the metabolic parameters approach that of WM.

Within a lesion ROI there exist voxels that are clearly WMLs (core) and those that are between WMLs and WM on the basis of T1 and FLAIR intensity. Thus, the observed metabolic parameter can be considered as $\widehat{y}={\chi }_c{y}_c+{\chi }_p{y}_p+\epsilon$ where χ_c = (1 − χ_p) = N Core Voxels/Total N Voxels in WML and y indicates the measured metabolic parameter in the core (c) and perilesional (p) areas. Within each subject, the multiple observations of $\widehat{y}$ at different thresholds allows for the solving for parameters y_c and y_p which are the core and perilesional box plots in Figure 2C. Unstable fits (R² < 0.75) were excluded from analysis resulting in two exclusions. These metabolic parameters were subjected to linear models with tissue (WM, GM, WML core, perilesional WM) as factor. Participant ID was included as a random effect. The numerical results of this analysis are shown in Figure 2 and the statistical results are shown in Supplemental Table 4.

Voxelwise Assessment of the Lesion – NAWM Boundary (Figure 3)

The preceding analysis assumes that metabolic parameters are constant across a ROI, in this case a tissue class. Relaxing that assumption requires estimating metabolic parameters voxelwise which has increased noise but higher spatial resolution. Region Based Voxelwise (RBV) partial volume correction uses GTM as an initial tissue class estimate and then incorporates spatial variability to yield a voxelwise partial volume corrected estimate²⁶. sGTM is numerically superior to GTM, thus we replaced GTM with sGTM in the RBV algorithm. We measured metabolic parameters parametric in distance from the lesion boundary using a lesion tissue score threshold of 0.7, the same threshold used in Figure 1. Single subject metabolic curves parametric in distance from lesion boundary was estimated using LOESS smoothing. A similar process was repeated parametric in distance from ventricles.

Comparison of Metabolic Features in Vulnerable vs. Non-vulnerable WM (Figure 4)

WMLs owing to SVD occur in a stereotypic pattern. To query if this topography was metabolically distinct, we separated WM into vulnerable (v) and non-vulnerable (n) WM. Lesion masks were transformed to a reference atlas space and the voxelwise frequency of lesions was compiled across the SVD group. This frequency map was then transformed back into each single subject’s anatomical space for further analysis. We next compared metabolic parameters in the healthy control groups in GM, WMv, and WMn. There is no a priori definition for a frequency cut off to define vulnerability to WMLs. Thus, we conducted this analysis over a range of lesion frequency thresholds to define vulnerable and non-vulnerable WM. At each threshold, sGTM was performed with WM separated by vulnerability status at that threshold.

Statistics and Reproducibility

This was an observational, cross-sectional study conducted on research volunteers without intervention. As such, group membership was not randomized. Investigators were not blind to group membership. No specific statistical method was used to pre-determine the sample size but our sample sizes are similar to other PET studies of metabolism. Analysis and statistical results were computed using MATLAB 2021 and R 4.2.1, respectively. The assumptions of linear model statistical models were met and the residuals were approximately normal as visualized by a Q-Q plot.

Data Availability

Processed data required to generate the figures and tables is included as supplemental information and at https://github.com/matthewbrier/AGinSVD. The Adult Metabolism & Brain Resilience (AMBR) dataset is in the process of being made availably for sharing to researchers with an appropriate Data Use Agreement due to Protected Health Information, limitations in the initial consent process, and regulations at the Knight Alzheimer Disease Research Center.