Petersen, Ronald C.

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.

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.

AbstractBackgroundStudies of aging and dementia often acquire brain MRI and PET images, then de‐identify and share these widely with other researchers. It is known that research participants can potentially be re‐identified from brain MRI and CT using face recognition, but this has not been tested with PET imaging.MethodWe generated face reconstruction images of 182 volunteers using their existing amyloid (PiB), tau (FTP), and FDG PET scans from clinical GE PET/CT scanners, and we measured how often commercial automatic face recognition (Microsoft Azure) correctly matched these reconstructions with photographs of their faces. We additionally compared these match rates with face reconstructions from CT (from PET/CT) and MRI. When testing face recognition from PET scans, CT was used only during PET reconstruction, not for face recognition. We also validated our updated mri_reface software, originally developed to de‐identify (“de‐face”) faces in MRI, with new support for PET images. We measured face recognition rates before and after applying mri_reface to PET, CT, and MRI. In addition, we measured effects of de‐facing on biomarker measurements (global and regional SUVR) using 244 ADNI Florbetapir (FBP) amyloid PET scans with two automated pipelines: FreeSurfer 6.0 (PETSurfer) and an in‐house SPM12‐based pipeline.ResultRates of successful face recognition with PET ranged from 32‐41% (Table 1) across tracers. These were lower than MRI (98%) and CT (78%) but still concerning for participant privacy. Using mri_reface reduced face recognition to 0‐3.5% for PET, 5% for CT, and 8% for MRI. mri_reface had little effect on global amyloid SUVR measurements from ADNI PET images (Figure 1). ICC values between original and de‐faced images were 1.00 with biases and median absolute differences < 0.5%, across both pipelines. Effects on individual regions all had ICC values > 0.98 with biases and median absolute differences < 2%.ConclusionFace reconstructions from PET images have sufficient quality for potential re‐identification using publicly available face recognition technology. Research studies should consider using face de‐identification (de‐facing) software on PET images in addition to CT and structural MRI. Our free, automatic mri_reface software greatly reduced the potential for face recognition without substantially affecting SUVR measurements.

AbstractBackgroundQuantitative susceptibility mapping (QSM) can detect iron distribution in the brain by estimating the local tissue magnetic susceptibility of every voxel. Dysregulation of neuronal iron has been correlated with typical Alzheimer’s disease (tAD), but little is known about iron distribution in atypical clinical presentations of AD such as logopenic progressive aphasia (LPA) and posterior cortical atrophy (PCA).MethodEleven PCA patients and eight LPA patients were recruited by the Neurodegenerative Research Group at Mayo Clinic, Rochester, MN, and underwent MRI that included a five‐echo gradient echo sequence for calculation of QSM. Mean QSM signal was extracted from gray and white matter for regions‐of‐interest across the brain using the Mayo Clinic Adult Lifespan Template. Hierarchical models were fit per‐region and per‐hemisphere to compare PCA and LPA to each other, to 20 healthy controls, and to 20 patients with tAD recruited by the Alzheimer’s Disease Research Center.ResultStrong evidence (posterior probability>0.99) was observed for greater susceptibility in the left middle occipital gyrus and amygdala in both LPA and PCA, and in the right precuneus and inferior temporal gyrus in PCA compared to controls. Moderate evidence for greater susceptibility (posterior probability>0.90) was also observed in the left insula, inferior occipital gyrus, entorhinal cortex, and right inferior parietal gyrus and amygdala in both LPA and PCA, along with right angular gyrus and middle occipital gyrus in PCA and right insula and caudate in LPA, when compared to controls. Comparing between phenotypes, LPA showed moderate evidence for greater susceptibility in the caudate, hippocampus, substantia nigra, left supramarginal gyrus and posterior cingulate compared to PCA, while PCA showed greater susceptibility in the right precuneus compared to LPA. Both LPA and PCA showed moderate‐strong evidence for greater susceptibility than tAD, particularly in medial and lateral parietal regions.ConclusionUnique iron profiles were identified in LPA and PCA within cortical and subcortical structures. These changes are in line with disease‐related neurodegenerative changes in the clinical phenotypes, suggesting that QSM could be an informative marker to study iron deposition in these patients.

AbstractBackgroundMayo Test Drive (MTD): Test Development through Rapid Iteration, Validation and Expansion, is a web‐based platform for self‐administered assessment for novel (Stricker Learning Span) and open access (Symbols Test) cognitive measures with previously demonstrated validity and reliability. Our study aim was to examine the usability of MTD, defined as the extent to which MTD was used to achieve the desired study goal of completing a testing session among participants willing to engage in an all‐remote study.Methods248 Mayo Clinic Study of Aging and Mayo Alzheimer’s Disease Research Center participants were invited to participate in this ancillary, uncompensated remote pilot study via email. To examine 1) willingness to participate in a remote cognitive assessment study and 2) usability of MTD, the initiation and completion rates of MTD were examined by cognitively unimpaired (CU; n = 142) and cognitively impaired (CI; n = 106) groups (see Figure 1). These rates and other factors associated with MTD feasibility were further explored across age in CU (i.e., 50‐69, 70‐79, and 80+ years) and specific diagnostic groups in CI (i.e., individuals with dementia, mild cognitive impairment (MCI), and possible MCI).ResultsSession completion rate (usability) was 98% across all participants who chose to participate, and, surprisingly, completion rates did not vary by age or diagnostic groups (p =.69; see Table 1). As expected, more individuals chose to participate in the CU (77%) versus CI (52%) group (p < .001). In terms of feasibility, individuals over 80 and with CI were more likely to have some contact with the study coordinator. Use of smartphones decreased with age, with most CU individuals over 70 using a personal computer (69‐73%; Table 2).ConclusionsThough the CI group was less likely to choose to participate than the CU group, out of those that did choose to participate the rate of completion was 98%. This demonstrates high usability of MTD. Study coordinator support and a flexible testing platform is critical for facilitating participation in remote testing across the aging population, and more supports and incentives may be needed for individuals with cognitive impairment to improve participation rates.

AbstractIntroductionAlzheimer's disease (AD) patients ≥70 years show smaller medial temporal volumes despite less 18F‐flortaucipir‐positron emission tomography (PET) uptake than younger counterparts. We investigated whether TAR DNA‐binding protein 43 (TDP‐43) was contributing to this volume‐uptake mismatch.MethodsSeventy‐seven participants with flortaucipir‐PET and volumetric magnetic resonance imaging underwent postmortem AD and TDP‐43 pathology assessments. Bivariate‐response linear regression estimated the effect of age and TDP‐43 pathology on volume and/or flortaucipir standardized uptake volume ratios of the hippocampus, amygdala, entorhinal, inferior temporal, and midfrontal cortices.ResultsOlder participants had lower hippocampal volumes and overall flortaucipir uptake. TDP‐43‐immunoreactivity correlated with reduced medial temporal volumes but was unrelated to flortaucipir uptake. TDP‐43 effect size was consistent across the age spectrum. However, at older ages, the cohort mean volumes moved toward those of TDP‐43‐positives, reflecting the increasing TDP‐43 pathology frequency with age.DiscussionTDP‐43 pathology is a relevant contributor driving the volume‐uptake mismatch in older AD participants.Highlights TDP‐43 pathology affects medial temporal volume loss but not tau radiotracer uptake. Greater TDP‐43 pathology effect is seen in old age due to its increasing frequency. TDP‐43 pathology is a relevant driver of the volume‐uptake mismatch in old AD patients.

ImportanceIndividuals with total joint arthroplasty (TJA) have long-term exposure to metal-containing implants; however, whether long-term exposure to artificial implants is associated with cognitive function is unknown.ObjectiveTo compare long-term cognitive trajectories in individuals with and without TJA.Design, Setting, and ParticipantsThis population-based cohort study assessed serial cognitive evaluations of 5550 participants (≥50 years of age) from the Mayo Clinic Study of Aging between November 1, 2004, and December 31, 2020.ExposuresTotal joint arthroplasty of the hip or the knee.Main Outcomes and MeasuresLinear mixed-effects models were used to compare the annualized rate of change in global and domain-specific cognitive scores in participants with and without TJA, adjusting for age, sex, educational level, apolipoprotein E ε4 carrier status, and cognitive test practice effects.ResultsA total of 5550 participants (mean [SD] age at baseline, 73.04 [10.02] years; 2830 [51.0%] male) were evaluated. A total of 952 participants had undergone at least 1 TJA of the hip (THA, n = 430) or the knee (TKA, n = 626) before or after entry into the cohort. Participants with TJA were older, more likely to be female, and had a higher body mass index than participants without TJA. No difference was observed in the rate of cognitive decline in participants with and without TJA until 80 years of age. A slightly faster cognitive decline at 80 years or older and more than 8 years from surgery was observed (b = −0.03; 95% CI, −0.04 to −0.02). In stratified analyses by surgery type, the faster decline was observed primarily among older participants with TKA (b = −0.04; 95% CI, −0.06 to −0.02).Conclusions and RelevanceIn this cohort study, long-term cognitive trajectories in individuals with and without TJA were largely similar except for a slightly faster decline among the oldest patients with TKA; however, the magnitude of difference was small and of unknown clinical significance.