Gunter, Jeffrey L.

AbstractBackgroundResting state functional connectivity (rsfMRI) studies have shown functional network disruptions in logopenic progressive aphasia (LPA) (reduced connectivity within the left temporal language network) and posterior cortical atrophy (PCA) (reduced connectivity within ventral visual network). However, little is known about how these networks compare and how connectivity differs within and between networks in these atypical Alzheimer’s Disease (AD) phenotypes.MethodFifty‐nine amyloid‐positive PCA patients and eighty‐five amyloid‐positive LPA patients were recruited by the Neurodegenerative Research Group at Mayo Clinic, Rochester, MN, and underwent structural and resting state functional MRI. Spatially preprocessed data was analyzed using the CONN functional connectivity toolbox to explore the default mode network (DMN), salience, sensorimotor, left language, visual and memory networks, which were generated by averaging multiple regions‐of‐interest (ROI) within each network. The data was analyzed at both the voxel‐level and network‐level (ROI‐ROI).ResultVoxel‐level differences in functional connectivity between PCA and LPA were observed in all networks (p<0.05; cluster size, FWE=0.05). PCA showed reduced connectivity in the occipital lobe in visual and memory networks, parietal lobe in language and DMN, and sensorimotor strip in sensorimotor network, particularly on the right, compared to LPA. Conversely, LPA showed reduced connectivity in the frontal cortex in salience, memory, visual and sensorimotor networks, inferior parietal lobe in visual network, and anterior temporal pole and sensorimotor cortex in the language network, compared to PCA. Between network differences (p<0.05; cluster size, FDR=0.05) were observed between groups, with PCA showing reduced connectivity between the visual and sensorimotor networks, while LPA showed reduced connectivity between the language and sensorimotor networks. The DMN was anti‐correlated to the language network in PCA, while it was positively correlated to the language network in LPA. In both phenotypes, the DMN was anticorrelated with the salience network. The visual network in PCA showed connections to all networks, while in LPA, it was only connected to the salience, language and DMN.ConclusionThe atypical AD phenotypes were associated with multiple disruptions in connectivity, both within and between brain networks. Phenotype specific differences in connectivity patterns were noted in visual network for PCA and language network for LPA.

AbstractBackgroundFluorodeoxyglucose positron emission tomography (FDG‐PET) is an established modality for differential diagnosis of dementia. Deriving phenotypic signatures from FDG‐PET via machine learning is challenging due to the high dimensional nature of brain images relative to the generally small number of labeled examples available for training, the class imbalance among those labels, and the cooccurrence of multiple pathologies. In this study, we developed a multi‐class, multi‐label framework to address these challenges.MethodA database of clinically acquired PET/CT images from 3,000 unique patients was used to develop a latent space model using matrix decomposition. This model was then applied to images from a separate cohort of Mayo Clinic Alzheimer’s Disease Research Center participants (n=1,745) labeled as cognitively unimpaired (CU) (n=1,436) or with the following potentially co‐occurring phenotypes: Alzheimer’s disease (AD) (n=165), Lewy body dementia (DLB) (n=92), behavioral variant frontotemporal dementia (bvFTD) (n=43), semantic (svPPA) (n=10) and logopenic (lvPPA) (n=13) variant PPA, and posterior cortical atrophy (PCA) (n=17). A k‐nearest neighbors classifier that is robust to these imbalanced and overlapping labels was then trained on these examples. The resulting classifier was evaluated by area under receiver‐operator characteristic curve (ROC‐AUC) via leave one out cross validation, using clinical diagnosis as the gold standard.ResultROC curves and AUC scores for each phenotype are illustrated in Fig. 1a. Because the classifier is based on a k‐nearest neighbors connectivity matrix, it has a convenient graphical representation, where images are nodes and edges are drawn between an image and its set of nearest neighbors in latent space. A self‐organizing force directed graph constructed in this way is illustrated in Fig 1b, highlighting the strong separation of CU and degenerative images, as well as the segregation of each phenotype within the neurodegenerative region of the graph.ConclusionIn this study, we developed a machine learning framework for classification of neurodegenerative disease based on k‐nearest neighbor analysis in a low dimensional latent space projection of FDG‐PET images. By leveraging low‐dimensional representations and k‐nearest neighbors analysis, this framework is robust in multi‐class, multi‐label tasks with strong class imbalance and provides a highly interpretable graphical representation.

AbstractBackgroundIn 2018 we presented the Computational DESH (CDESH) metric (DOI:10.1016/j.nicl.2018.11.015), a machine learning measure of abnormal CSF space patterns commonly observed on MRI in Disproportionately Enlarged Subarachnoid‐space Hydrocephalus (DESH‐type) NPH. We have shown that CDESH positive individuals often have discordant CSF and PET amyloid measures (DOI:10.1016/j.jalz.2019.06.4921). Positive CDESH scores are predictive of cognitive decline (DOI:10.1016/j.neuroimage.2021.117899, DOI: 10.1212/WNL.0000000000008616). The score is calculated from regional CSF volumes in a sulcal atlas and within the ventricles using a support vector machine (SVM). The Mayo Clinic Study of Aging (MCSA) changed scanner platforms (GE to Siemens) and T1‐weighted image resolution (1.2mm x 1.05mm x 1.05mm to 0.8mm isotropic) in 2017. We refer to these as “isotropic” or “anisotropic” scans. Our objective was to assess the dependence of a machine learning method (specifically CDESH) to image input variation and, if necessary, determine what needed corrections.MethodAll data considered is from the MCSA or Mayo Clinic ADRC. The SVM was trained on anisotropic images with laboriously labelled ground truth. We resampled the isotropic data from a crossover study (N=112, typically scanned same‐day with both platforms/resolutions) to match the anisotropic resolution. Two approaches to correction were evaluated (Figure 1). In Figure 2 we compare of the score with and without correction cross‐sectionally in 998 MCSA participants. In Figure 3 we present longitudinal series from 23 participants with multiple scans using the isotropic resolution protocol and at least one prior anisotropic image.ResultsComparisons of data from the crossover study are presented in Figure 1. Matching voxel resolution to the data used to train the SVM is critical and a subsequent mild linear score adjustment may be warranted. Cross‐sectional comparisons (Figure 2) confirm improvement matching voxel resolution with a possible mild linear score adjustment. After matching voxel resolution CDESH score trajectories (Figure 3) are markedly more stable across this scanner/protocol change.ConclusionThe SVM‐based CDESH score is sensitive to voxel resolution. Corrections are needed in image processing and possibly in the scalar output score to achieve stability across protocol and scanner changes. The resulting correction supports coherent long timeseries data and longitudinal extension of previous work.

AbstractBackgroundCerebral Microbleeds (CMBs) are tracked in anti‐amyloid drug safety monitoring. T2*GRE or SWI MRI images are visually assessed for the presence of CMBs. Automated methods tuned for sensitivity often have unacceptably high false positive rates requiring extensive human cleanup.MethodAssume contemporaneous T2*GRE and T1‐weighted images. CMB detection (Figures 1,2) has three steps: 1) Segmentation outputs from the T1‐weighted image are resampled into the T2*GRE image voxel raster; 2) Normalized cross‐correlation of the T2*GRE with a CMB template identifies “candidate CMBs”; 3) a neural network using voxel patches centered on each candidate CMB rejects false positives. Patches presented to the network are 4D and include five contrasts drawn from the T2*GRE, correlation, T1‐weighted, WM probability and CSF probability images. The simple network architecture consists of convolutional blocks followed by dense blocks. The system is trained using images from ADNI, ARIC, MCSA, and the Mayo Clinic ADRC that have been previously visually graded for CMBs (Table in Figure 3). Importantly, we include participants with and without CMBs in our data reflecting the fact that the majority of people (even the elderly) have no CMBs. Data is split 70:10:20 by subject for training, validation, and testing. Various network configurations were evaluated.ResultHigh candidate detection sensitivity can be achieved at the expense of presenting hundreds to thousands of false positive candidates per input image to the rejection network. Our best performing network configuration to date has AUC 0.982 (95% bootstrap CI 0.971 to 0.990 based on the 10% validation subset). Choosing an operating point with 95% sensitivity in the validation set and admits, on average, 0.6 false positive CMB candidates per input T2*GRE image. Sensitivity is hindered by a small number of obvious errors in the ground truth visual assessment. Validation loss tracks training loss with no indication of over‐fitting. Example patches are shown in Figure 4.ConclusionExcessive false positive detection requires significant human correction and effort, preventing automated CMB detection from being widely adopted. We have an automated detection system that is sensitive (95%) and provides excellent false positive rejection (<1 false positive CMB/input image) reducing needed human effort.

AbstractBackgroundAlzheimer’s disease (AD) pathologic changes (e.g. amyloid and tau aggregation), and cerebrovascular disease often coexist, and it remains unclear whether these are mechanistically linked processes (Graff‐Radford, Brain 2019; Provenzano, JAMA Neurol 2013). Most studies evaluating these associations are cross‐sectional; however, establishing temporal relationships may provide additional insight into potential mechanistic associations. One method to assess the temporal evolution of a biomarker is the Accelerated Failure Time (AFT) model, which assumes a common pattern of progression for a given biomarker that is shifted earlier or later in time for each individual (Therneau, Neuroimage 2021). In this work we investigate the relationship of disease trajectories of cerebrovascular disease (via white matter hyperintensity, WMH, volume) and AD biomarkers (via amyloid and tau PET).MethodsWe included Mayo Clinic Study of Aging (MCSA) and Mayo Alzheimer’s Disease Research Center (ADRC) participants with amyloid PET, tau PET, and MRI, age > 50 years, and a clinical diagnosis of cognitively unimpaired (CU), mild cognitive impairment (MCI) or Alzheimer’s disease dementia. The AFT model was fit with amyloid PET (PiB) global meta‐ROI SUVR, tau PET (flortaucipir) temporal meta‐ROI SUVR, and WMH volume as the tri‐variate endpoints, with APOE genotype, sex, education, total intracranial volume, and referral/population subject as covariates. A per subject/endpoint random effect represents individual adjustments for each biomarker to indicate if the subject's biomarker became abnormal earlier or later relative to their demographic peers, accounting for the covariates.ResultsThe study included 1144 participants (Table 1). There was a strong correlation between earlier onset of amyloid and earlier onset of tau (r=0.64, CI=0.60‐0.67) (Figure 1, Table 2); early/late amyloid timing explained nearly half (0.642 = 41%) of the variation in tau onset. There were only weak associations between an individual’s WMH and amyloid PET onset adjustments (r=0.10, 1.0%) and WMH and tau PET onset adjustments (r=‐0.07, 0.5%).ConclusionThis analysis supports a strong mechanistic relationship between amyloid and tau aggregation, as reported by others. However, our findings do not support a mechanistic association between WMH and amyloid or tau PET.

AbstractBackgroundIn vivo tau‐positron emission tomography (PET) is an attractive biomarker for Alzheimer’s disease (AD) diagnosis and treatment. However, tau‐PET is less widely available than other modalities. In this study, we tested cross‐modality synthesis of tau‐PET brain images from fluorodeoxyglucose F‐18 (FDG)‐PET using a deep convolutional neural network (CNN).MethodParticipants (n=1,192) who had brain FDG‐PET with 18F‐FDG and tau‐PET with Flortaucipir (F‐18‐AV‐1451) were included for training and testing. This cohort spanned normal aging (ages 26‐98), pre‐clinical, and clinical AD and related disorders including the FTD and DLB spectrum. External validation was done using ADNI (n=288). The PET scans were co‐registered to the corresponding MRI and subsequently warped to Mayo Clinic Adult Lifespan Template (MCALT) space. Tau‐PET images were SUVR‐normalized to the cerebellar crus, and FDG to the pons. A 3D dense‐U‐net model was utilized as an architecture. Cross‐validation experiments were conducted using 5‐fold validations (60% training set, 20% validation set, and 20% test set) with mean squared error as the loss function.ResultOur dense‐U‐net model successfully synthesized tau‐PET from metabolic images with good correlation and low prediction error for regional SUVRs (Figure 1A‐C). The model showed a robust prediction ability, performing accurately in an independent, external ADNI cohort (Figure 1D‐F). The model‐imputed tau‐PET significantly improved performance in classifying tau positivity (mean AUROC(±SD)=0.78±0.04 and 0.85±0.03 for FDG‐PET and synthesized tau‐PET, respectively) and diagnostic groups (cognitively unimpaired with abnormal amyloid‐PET vs. cognitively impaired with abnormal amyloid‐PET) compared to the original input FDG data (mean AUROC(±SD)=0.89±0.04, 0.85±0.05 0.91±0.04 for actual tau‐PET, FDG‐PET and synthesized tau‐PET, respectively) (Figure 2), suggesting enhanced clinical utility for metabolic images. The ADNI cohort also showed similar results (for tau positivity: AUROC=0.66 and 0.78 for FDG‐PET and synthesized tau‐PET, respectively; for CU A+ vs. CI A+: AUROC=0.86, 0.62, and 0.73 for actual tau‐PET, FDG‐PET, and synthesized tau‐PET; Figure 3).ConclusionWe showed that using a CNN model to predict tau‐PET from FDG‐PET is feasible. The synthesized tau‐PET can augment the value of FDG‐PET, facilitating the multi‐modal diagnosis of AD.

AbstractBackgroundDisruption of the default mode network (DMN) is thought to be a common feature of aging and across Alzheimer’s disease (AD) clinical phenotypes, while syndromic variability could be reflected in selective degeneration of macro‐scale networks optimally supporting specific mental abilities (Fig 1). As part of a collaborative effort between Mayo Clinic and the Human Connectome Project‐Aging (HCP‐A), we aimed to delineate a biomarker of global network failure across aging and the dysexecutive (dAD) and amnestic phenotypes of AD, and biomarkers of focal degeneration specific to each AD phenotype.MethodWe used a regression framework to assess the relationships between a marker of DMN dysfunction, the network failure quotient (NFQ), and factors relevant to AD such as age, cortical thickness (global cortical thickness, inferior parietal lobe), hippocampal volume, and cognition (episodic memory, cognitive flexibility) in the HCP‐A cohort, which consists of 725 non‐demented older adults. We then fit centile curves to the HCP‐A cohort to assess whether the NFQ and focal degeneration markers could distinguish dAD (n=10) and amnestic AD (n=8) patients from the HCP‐A cohort at the group‐ and individual‐level. Participants were scanned using the HCP‐A protocol, allowing for the acquisition of high‐resolution multi‐echo structural images and longer resting‐state connectivity acquisition time compared to traditional protocols.ResultThe NFQ related to age, cortical thickness, hippocampal volume, and cognition in the HCP‐A cohort. It also strongly discriminated dAD and amnestic AD from the HCP‐A cohort but did not discriminate between AD phenotypes. In contrast, markers of focal degeneration were phenotype‐specific; parieto‐frontal areas yielded largest effect sizes for dAD patients, and hippocampal and temporal areas best discriminated patients with amnestic AD (Fig 2).ConclusionThe NFQ may reflect shared system‐level pathophysiological mechanisms across aging and the dysexecutive and amnestic phenotypes of AD whereas focal degeneration markers highlight distinct pathognomonic processes across AD phenotypes. These findings suggest that inter‐individual variability in cognitive impairment in AD would be closely tied to modular, phenotype‐specific network degeneration combined with a common global functional disruption of the DMN. Such information has crucial implications for clinical biomarkers and to provide a deeper understanding of heterogeneity in AD.

AbstractBackgroundData heterogeneity due to cross‐scanner variations is a major challenge in multi‐center aging and dementia studies. Brain structural measures for the same participant can vary due to scanner differences (manufacturer, scanner age/technology, signal‐to‐noise ratio, pulse sequence design, contrast, resolution) and result in biomarker variability. In this work, we applied basic image processing and deep learning methods (style transfer (ST) and generative adversarial networks (GANs)) for harmonizing T1‐weighted MRI scans across change in vendors (GE to Siemens) and resolution (1x1x1.2mm to 0.8x0.8x0.8mm) and test whether this harmonization improved compatibility of AD biomarker measures.MethodWe utilized training (n=814) and test (n=113 same participants scanned on two scanners close‐in‐time, typically on the same day, with different resolution) data from cognitively impaired and unimpaired participants scanned on GE and Siemens scanners from Mayo Clinic Study of Aging and Mayo ADRC. The harmonization was conducted to translate GE to Siemens. All scans were affine transformed to the same template space, resampled to 1.5 mm3, and bias corrected. Histogram matching was performed using cumulative density functions. A ResNet18 was trained for the ST’s content and style layers. Two types of GANs, conditional GANs (cGAN) and cycle‐consistent GANs (cycleGAN), were constructed. We then measured regional cortical thickness from each image using FreeSurfer‐7.1.1 and compared the thickness measurements from the original and translated scans using intraclass‐correlation (ICC) and Deming regression.ResultIn the test dataset, ST and cycleGAN showed the best performances when comparing the translated GE to Siemens. With ST, the ICC increased (vs. the original scans) slightly for the temporal meta‐ROI (0.956 from 0.954) and more substantially for the frontal pole ROI (0.706 from 0.594). The cycleGAN showed an increase in ICC for the frontal pole ROI (0.610 from 0.594) but not the meta‐ROI (Fig. 1).ConclusionEach method leveraged a different strength. Histogram matching captures contrast, ST extracts textural information, and GANs learn combination of contrast, edge and texture properties. We found that deep learning is a promising approach to cross‐scanner variability. Our comparisons revealed anatomical differences across methods; thus, disease specific measure should be considered for selection of the scanner harmonization method.

AbstractBackgroundCerebral Microbleeds (CMBs) are tracked in anti‐amyloid drug safety monitoring. T2*GRE or SWI MRI images are visually assessed for the presence of CMBs. Automated methods tuned for sensitivity often have unacceptably high false positive rates requiring extensive human cleanup.MethodAssume contemporaneous T2*GRE and T1‐weighted images. CMB detection (Figures 1,2) has three steps: 1) Segmentation outputs from the T1‐weighted image are resampled into the T2*GRE image voxel raster; 2) Normalized cross‐correlation of the T2*GRE with a CMB template identifies “candidate CMBs”; 3) a neural network using voxel patches centered on each candidate CMB rejects false positives. Patches presented to the network are 4D and include five contrasts drawn from the T2*GRE, correlation, T1‐weighted, WM probability and CSF probability images. The simple network architecture consists of convolutional blocks followed by dense blocks. The system is trained using images from ADNI, ARIC, MCSA, and the Mayo Clinic ADRC that have been previously visually graded for CMBs (Table in Figure 3). Importantly, we include participants with and without CMBs in our data reflecting the fact that the majority of people (even the elderly) have no CMBs. Data is split 70:10:20 by subject for training, validation, and testing. Various network configurations were evaluated.ResultHigh candidate detection sensitivity can be achieved at the expense of presenting hundreds to thousands of false positive candidates per input image to the rejection network. Our best performing network configuration to date has AUC 0.982 (95% bootstrap CI 0.971 to 0.990 based on the 10% validation subset). Choosing an operating point with 95% sensitivity in the validation set and admits, on average, 0.6 false positive CMB candidates per input T2*GRE image. Sensitivity is hindered by a small number of obvious errors in the ground truth visual assessment. Validation loss tracks training loss with no indication of over‐fitting. Example patches are shown in Figure 4.ConclusionExcessive false positive detection requires significant human correction and effort, preventing automated CMB detection from being widely adopted. We have an automated detection system that is sensitive (95%) and provides excellent false positive rejection (<1 false positive CMB/input image) reducing needed human effort.

AbstractBackgroundIn 2018 we presented the Computational DESH (CDESH) metric (DOI:10.1016/j.nicl.2018.11.015), a machine learning measure of abnormal CSF space patterns commonly observed on MRI in Disproportionately Enlarged Subarachnoid‐space Hydrocephalus (DESH‐type) NPH. We have shown that CDESH positive individuals often have discordant CSF and PET amyloid measures (DOI:10.1016/j.jalz.2019.06.4921). Positive CDESH scores are predictive of cognitive decline (DOI:10.1016/j.neuroimage.2021.117899, DOI: 10.1212/WNL.0000000000008616). The score is calculated from regional CSF volumes in a sulcal atlas and within the ventricles using a support vector machine (SVM). The Mayo Clinic Study of Aging (MCSA) changed scanner platforms (GE to Siemens) and T1‐weighted image resolution (1.2mm x 1.05mm x 1.05mm to 0.8mm isotropic) in 2017. We refer to these as “isotropic” or “anisotropic” scans. Our objective was to assess the dependence of a machine learning method (specifically CDESH) to image input variation and, if necessary, determine what needed corrections.MethodAll data considered is from the MCSA or Mayo Clinic ADRC. The SVM was trained on anisotropic images with laboriously labelled ground truth. We resampled the isotropic data from a crossover study (N=112, typically scanned same‐day with both platforms/resolutions) to match the anisotropic resolution. Two approaches to correction were evaluated (Figure 1). In Figure 2 we compare of the score with and without correction cross‐sectionally in 998 MCSA participants. In Figure 3 we present longitudinal series from 23 participants with multiple scans using the isotropic resolution protocol and at least one prior anisotropic image.ResultsComparisons of data from the crossover study are presented in Figure 1. Matching voxel resolution to the data used to train the SVM is critical and a subsequent mild linear score adjustment may be warranted. Cross‐sectional comparisons (Figure 2) confirm improvement matching voxel resolution with a possible mild linear score adjustment. After matching voxel resolution CDESH score trajectories (Figure 3) are markedly more stable across this scanner/protocol change.ConclusionThe SVM‐based CDESH score is sensitive to voxel resolution. Corrections are needed in image processing and possibly in the scalar output score to achieve stability across protocol and scanner changes. The resulting correction supports coherent long timeseries data and longitudinal extension of previous work.