Gebre, Robel K

Abstract Background Neuroimaging data typically arrives as Digital Imaging and Communications in Medicine (DICOM) files, batched by exam. We use exam or study to denote a single imaging session of a participant, containing up to ∼40 series (the generic term for scans and derived products). Different series have different properties and applications, making their labeling complex. Filenames often change in transit, and Series Description tags (0008, 103E) are arbitrary, operator‐entered, and unreliably coupled to the series type. Furthermore, label preferences vary from site to site, making the application of a common classification system mandatory for multisite studies. Here we describe the system that we have developed and applied to the Alzheimer's Disease Neuroimaging Initiative (ADNI) and SCAN. Method Although Series Description is unreliable, DICOM includes other tags that parameterize MR pulse sequences and suggest their purpose. The mapping from these features to types is difficult to explicitly specify, especially since different manufacturers and protocols use different conventions. Thus we developed a machine learning classifier using 4060 semiautomatically labeled head and spinal series from 1.5 to 7T MR scanners, including both site‐specific and standard protocols designed to study a wide range of neurological conditions, including dementia, hydrocephalus, and multiple sclerosis. An 80%/20% split, stratified by type, was used for the training/test sets. After initial training, a few cases in the training set with high errors due to rarity were augmented and the model retrained. Performance was evaluated on the clean test set. Result The taxonomy is presented in Table 1. The classifier correctly identified 807 out of 812 (99.4%) series, with the error cases coming from outside the ADNI or SCAN protocols (Table 2). The confusion matrix is shown in Figure 1. Conclusion We implemented an automatic series type classifier with labels chosen to be generally useful for the brain MRI community. Some errors were noted, but the test set was deliberately challenging and the errors appear to be ones that could easily be made by a human. The classifier is expected to have good tolerance for as yet unseen head MR protocols but is likely unsatisfactory for body MRI.

Abstract Background Cognitive resilience to Alzheimer's disease (AD) pathology has been widely studied, but distinguishing resilient individuals from those with typical tau pathologies remains challenging due to the overlap in imaging and plasma biomarker profiles. We hypothesize that our machine learning‐based (ML) tau heterogeneity score, THETA (JNUMED, July 2024), could serve as an effective marker of resilience and provide a better estimation of tau's temporal progression. Method Longitudinal imaging data from the Mayo Clinic Study of Aging (MCSA, n = 1834) and ADRC ( n = 246) were analyzed. Participants were categorized into typical AD and resilient groups using amyloid PiB‐PET (abnormal PiB‐PET≥25 Centiloid units), Flortaucipir (FTP‐PET) (meta‐ROI≥1.29 SUVr, ERC≥1.27 SUVr) and Mini‐Mental State Examination (MMSE). Typical AD was defined as abnormal PiB‐PET and abnormal FTP‐PET for both biomarkers (Figure 1). Resilience to AD pathology was determined by predicted values of MMSE >28 from a linear mixed effects model for cognitively unimpaired participants. In addition, a resistant group was broadly defined as having normal PiB‐PET. To capture the spatial heterogeneity of tau distribution in the brain, regional THETA scores were generated using ML model and visual ratings. Additionally, a progression model, Sampled Iterative Local Approximation (SILA), was fitted using AD and plasma biomarkers ( p ‐tau181 and GFAP) for validation. Result THETA identified regions of reduced tau uptake in the frontal, parietal, temporal, occipital, and limbic regions when compared to the typical AD group (Figure 2A). This was confirmed by the preserved cortical thickness in these same regions (Figure 2D). In addition, SILA showed THETA can distinguish global tau temporal progressions of typical AD, resilient, and resistant groups compared to the other biomarkers (Figure 3). Interestingly, the onset age for GFAP was younger for resilient and typical AD, indicating earlier onset of inflammation in the presence of pathology. Conclusion The significant regions identified by THETA, showing reduced tau and preserved cortical thickness, combined with the clear offset in progression models between resilient and typical AD, suggest that THETA may help uncover pathological tau mechanisms that differentiate these groups. These findings highlight the potential of THETA to provide valuable insights into resilience mechanisms and enhance our understanding of AD.

Abstract Background Brain age models are increasingly being used as measures of brain health and neurodegeneration. These models can be deep learning (DL) based and/or classical methods and provide a brain age gap (BAG) which is the difference between chronological and predicted brain age. Our goal was to develop DL‐based scanner invariant brain age models that can be applied in aging and dementia studies even when there are scanner changes that may hinder longitudinal follow‐up. Method We identified 3374 cognitively unimpaired (CU) participants with T1 scans from the Mayo Clinic Study of Aging (MCSA). A subset of 863 CU amyloidosis negative (PiB‐PET cutoff < 25 Centiloid) and no history of diabetes and hypertension was selected (CU‐noVAβ). We built two models – Model‐A trained on CU‐noVAβ, and Model‐B trained on all CU. We included 453 MCSA MCI participants as test set for both models. We used 113 participants (74 CU, 28 MCI, 10 dementia) from MCSA and Mayo ADRC with scans from two different vendors (scanned within a week of each other) for testing cross‐vendor/scanner differences (cross‐vendor data). The two unique features of our model included histogram matching and scanner type as an input to the DenseNet model for training (Figure 1). Result In the data, the inclusion of scanner as an input decreased the prediction of mean age differences between the scanners (Model‐A=2.17 years; Model‐B=1.71 years) (Figure 2). The mean absolute error (MAE=absolute value of BAG) for Model‐A in MCI was 5.72 years and Model‐B in MCI was 4.06 years, consistent across scanners (Figure 3). Longitudinal prediction increased with age (Model‐A slope=0.53; Model‐B slope=0.63). Conclusion We found that training on all CU (Model B) rather than a subset CU‐noVAβ (Model A) allowed for the DenseNet model to backpropagate on more variable data which lowered MAE in prediction error between scanners and lowered MAE for MCIs. Histogram matching and scanner as an input to the brain age model accounted for scanner changes and resulted in scanner invariant longitudinal brain age estimation.

Abstract Background Multi‐site aging and dementia studies are challenged by data heterogeneity caused by scanner (manufacturer and protocol) related variability. Quantitative analysis of white matter hyperintensity (WMH) from pooled samples, a traditional marker of cerebrovascular disease, is sensitive to machine variability. Our goal was to compare Deming (symmetric) linear regression and ComBat, a common method used for harmonizing neuroimaging measurements, for harmonizing WMH measurements on participants scanned on two different scanners (GE and Siemens) and protocols (2D FLAIR on GE and 3D FLAIR on Siemens). Methods Cognitively unimpaired (CU) ( n =53, age (mean (SD))=67 (15) years) and cognitively impaired (CI; MCI/dementia) ( n =34, age (mean (SD))=82 (9) years) participants from the Mayo Clinic Study of Aging and Mayo ADRC were scanned on GE and Siemens scanners within a week of each other (typically done on the same day). WMH as a percentage of total intracranial volume was measured using multispectral segmentation of FLAIR and T1‐weighted images (Gunter et. al. AAIC 2023). We used Deming regression as it includes errors in both input dimensions. ComBat was applied using age, sex, and diagnosis as covariates, with scanner vendor and protocol as the batch variable. We assessed the inter‐scanner similarity using Maximum Mean Discrepancy (MMD) analysis, where a value of zero indicates identical distributions, p >0.05 indicates no significant difference under H 0 . (Gretton et al., JMLR 2012). Results Our results showed that the distribution of WMH measurements, which were different in the CU group, were more similar after Deming than ComBat harmonization (Figure 1). This was confirmed by the lower MMD in the CU group from Deming regression (‐0.01 ( p =0.76) from 0.03 ( p =0.05)) compared to ComBat (0.04 ( p =0.02))(Figure 1). However, ComBat and Deming were unable to create distributions as compatible as original measurements for the CI group. Conclusion Using a unique dataset of the same participants scanned close in time on different scanner and protocols, we found that (1) MMD measurements were useful to test the need/improvement due to harmonization; (2) ComBat was not substantially better at creating compatible distributions between scanner vendors than Deming regression, and both were sometimes inadequate. Future work will require alternative models to linear scale and shift transformations.

Abstract Background Neuroimaging data typically arrives as Digital Imaging and Communications in Medicine (DICOM) files, batched by exam. We use exam or study to denote a single imaging session of a participant, containing up to ∼40 series (the generic term for scans and derived products). Different series have different properties and applications, making their labeling complex. Filenames often change in transit, and Series Description tags (0008, 103E) are arbitrary, operator‐entered, and unreliably coupled to the series type. Furthermore, label preferences vary from site to site, making the application of a common classification system mandatory for multisite studies. Here we describe the system that we have developed and applied to the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and SCAN. Method Although Series Description is unreliable, DICOM includes other tags that parameterize MR pulse sequences and suggest their purpose. The mapping from these features to types is difficult to explicitly specify, especially since different manufacturers and protocols use different conventions. Thus we developed a machine learning classifier using 4060 semiautomatically labeled head and spinal series from 1.5 to 7T MR scanners, including both site‐specific and standard protocols designed to study a wide range of neurological conditions, including dementia, hydrocephalus, and multiple sclerosis. An 80%/20% split, stratified by type, was used for the training/test sets. After initial training, a few cases in the training set with high errors due to rarity were augmented and the model retrained. Performance was evaluated on the clean test set. Result The taxonomy is presented in Table 1. The classifier correctly identified 807 out of 812 (99.4%) series, with the error cases coming from outside the ADNI or SCAN protocols (Table 2). The confusion matrix is shown in Figure 1. Conclusion We implemented an automatic series type classifier with labels chosen to be generally useful for the brain MRI community. Some errors were noted, but the test set was deliberately challenging and the errors appear to be ones that could easily be made by a human. The classifier is expected to have good tolerance for as yet unseen head MR protocols but is likely unsatisfactory for body MRI.

AbstractBackgroundChronic medical conditions increase dementia risk, and the mechanism may be captured computationally by measures of accelerated brain aging. We hypothesized that quantifying the impact of chronic conditions on GM and WM aging will aid in measuring the differential impact of each condition on brain aging and shed light on the mechanisms through which they increase dementia risk. We specifically focused on microstructural brain changes with Neurite Orientation Dispersion and Density Imaging (NODDI) based on advanced diffusion MRI ‐ neurite density index (NDI), free water fraction (FWF), and orientation dispersion index (ODI).MethodWe identified 1123 participants [913 cognitively unimpaired (CU), 114 with MCI, 96 with AD] from the Mayo Clinic Study of Aging (MCSA) and Mayo ADRC with multi‐shell diffusion MRI. MCSA participants had chronic medical conditions’ diagnosis (based on US Dept of HHS) extracted from the electronic health care records. Using the BorutaSHAP algorithm on NDI, ODI, and ISOVF features from GM and WM separately, we constructed two distinct brain age models for GM and WM. First, we evaluated the usefulness of the predicted age gap (PAG), which is the difference between predicted age by model and chronological age, for both WM and GM NODDI models in CU/MCI/AD. We then developed separate GM and WM models for each condition using samples from condition‐negative participants to compare PAG of GM and WM for that condition.ResultThere was greater PAG in GM (avg = ) than WM (avg = 9.15 yrs) in MCI/AD participants (Fig. 1A). Each condition shifted the GM PAG by 0.10 to 3.56 years and WM PAG by 0.38 to 3.72 years with stroke and congestive heart failure showing the greater PAG overall for GM and WM PAD respectively (Fig. 2). COPD, stroke, chronic kidney disease, hypertension had comparatively greater GM PAG than WM PAG; while hyperlipidemia, congestive heart failure and coronary artery disease had greater WM PAG than GM PAG (Fig. 1B).ConclusionChronic conditions are associated with increased GM and WM Brain aging models can aid in quantifying the impact of chronic conditions on dementia risk and understanding mechanisms.

AbstractBackgroundBiomarker cutpoints have received significant importance in AD research. Reliable worsening method based on serial biomarker data has been proposed (Jack et al., 2017 Alz & Dem). Advanced machine learning models (ML) and explainable AI provide an opportunity to enhance existing techniques. We aimed to estimate biomarker cutpoints on cross‐sectional data using ML by identifying important transitions in the biomarker values.MethodWe first trained an ML model with the biomarker of interest as an input including other available biomarkers, demographics, and diagnosis, and predicted global cognition z‐score. We then obtained SHAP values (Shapley Additive exPlanations), fit a curve between the SHAP values and biomarker values, and calculated the first derivatives (slope) of the curve. Then, a CUSUM (Cumulative Sum) analysis was used to identify SHAP change points and cross‐referenced to values where the slopes were steepest near the zero‐crossing location of the curve. A zero SHAP is an important indicator of a feature’s progressive importance for model prediction because a feature’s influence on prediction significantly shifts when its SHAP values go from positive to negative or vice versa. However, relying on zero‐crossing alone may be insufficient when the relationship is non‐linear. Consequently, we calculated the second derivatives to identify inflection points (peaks and troughs) along the curve. Lastly, we compiled all the potential cutpoints and performed post‐processing as shown in Fig. 1.ResultThe algorithm identified cutpoints at 1.44 SUVR for all and at 1.42 SUVR for CU participants (Fig. 2). At age 70, both cutpoints estimated the percentage of abnormals at less than 40% (Fig. 3).ConclusionTPE amyloid‐PET SUVR cutpoints for CUs equal the reported RW values (1.42 SUVR). A strength of the TPE is its application on cross‐sectional data making it versatile for various biomarkers across different cohorts.

AbstractBackgroundChronic medical conditions increase dementia risk, and the mechanism may be captured computationally by measures of accelerated brain aging. We hypothesized that quantifying the impact of chronic conditions on GM and WM aging will aid in measuring the differential impact of each condition on brain aging and shed light on the mechanisms through which they increase dementia risk. We specifically focused on microstructural brain changes with Neurite Orientation Dispersion and Density Imaging (NODDI) based on advanced diffusion MRI ‐ neurite density index (NDI), free water fraction (FWF), and orientation dispersion index (ODI).MethodWe identified 1123 participants [913 cognitively unimpaired (CU), 114 with MCI, 96 with AD] from the Mayo Clinic Study of Aging (MCSA) and Mayo ADRC with multi‐shell diffusion MRI. MCSA participants had chronic medical conditions’ diagnosis (based on US Dept of HHS) extracted from the electronic health care records. Using the BorutaSHAP algorithm on NDI, ODI, and ISOVF features from GM and WM separately, we constructed two distinct brain age models for GM and WM. First, we evaluated the usefulness of the predicted age gap (PAG), which is the difference between predicted age by model and chronological age, for both WM and GM NODDI models in CU/MCI/AD. We then developed separate GM and WM models for each condition using samples from condition‐negative participants to compare PAG of GM and WM for that condition.ResultThere was greater PAG in GM (avg = ) than WM (avg = 9.15 yrs) in MCI/AD participants (Fig. 1A). Each condition shifted the GM PAG by 0.10 to 3.56 years and WM PAG by 0.38 to 3.72 years with stroke and congestive heart failure showing the greater PAG overall for GM and WM PAD respectively (Fig. 2). COPD, stroke, chronic kidney disease, hypertension had comparatively greater GM PAG than WM PAG; while hyperlipidemia, congestive heart failure and coronary artery disease had greater WM PAG than GM PAG (Fig. 1B).ConclusionChronic conditions are associated with increased GM and WM Brain aging models can aid in quantifying the impact of chronic conditions on dementia risk and understanding mechanisms.

AbstractBackgroundWe recently developed a novel tau‐PET summary measure THETA, capturing regional heterogeneity and identifying tau status, using ground truth visual assessments from a large single‐center cross‐sectional dataset and validated on independent cohorts [1, 2]. In this study, we aimed to evaluate the performance of THETA on longitudinal and histopathology data.MethodWe included longitudinal tau‐PET ([18F]flortaucipir) data from 696 Mayo Clinic Study of Aging (MCSA) and ADRC participants, with histopathology in n = 90. Fig. 1 shows the model that uses regional standard uptake value ratios (SUVR) and a target of binary class of tau positivity for prediction of THETA. This model was applied to predict tau status on each followup scan. Slopes of the meta‐ROIs’ tau SUVRs and THETA were evaluated by diagnostic group and as a function of baseline amyloid SUVR in discordant CU participants (where meta‐ROI and visual assessments did not match at baseline) and in incident‐THETA+ CU participants (whose THETA moved from below to above 1 over serial tau‐PET scans). We evaluated tau measurements as a function of Braak stages for neurofibrillary tangles.ResultLongitudinal plots of meta‐ROIs for diagnostic groups were very similar, but the expanded range of THETA based on visual positivity prediction clearly identified tau positive or tau‐negative scans. In discordant‐CU (n = 97) and incident‐THETA+ CU participants (n = 14 all amyloid positive at follow‐up but only 65% at baseline), the relationship between baseline amyloid and rate of tau increase was stronger for THETA than temporal meta‐ROI (Fig. 2). Separation between baseline and follow‐up was greater for THETA (t‐statistics = 90) compared to temporal meta‐ROI (t‐statistics = 20) (p<0.01) in the incident‐THETA+ CU participants. THETA showed a slightly stronger association with Braak stage than meta‐ROIs (rho = 0.87 vs. ≤0.83, p<0.05), with better separation of clinical diagnoses (Fig. 3).ConclusionTHETA remained clearly negative or positive in MCI and AD, providing consistent information on underlying etiology of impairment at both baseline and follow‐up. Although binary in its construction, THETA both provided separation of values based on tau status and its change correlated with baseline amyloid burden especially in discordant‐CU where tau deposition is not in typical meta‐ROIs. Further work is needed to confirm if THETA captures early tau changes.

AbstractBackgroundBiomarker cutpoints have received significant importance in AD research. Reliable worsening method based on serial biomarker data has been proposed (Jack et al., 2017 Alz & Dem). Advanced machine learning models (ML) and explainable AI provide an opportunity to enhance existing techniques. We aimed to estimate biomarker cutpoints on cross‐sectional data using ML by identifying important transitions in the biomarker values.MethodWe first trained an ML model with the biomarker of interest as an input including other available biomarkers, demographics, and diagnosis, and predicted global cognition z‐score. We then obtained SHAP values (Shapley Additive exPlanations), fit a curve between the SHAP values and biomarker values, and calculated the first derivatives (slope) of the curve. Then, a CUSUM (Cumulative Sum) analysis was used to identify SHAP change points and cross‐referenced to values where the slopes were steepest near the zero‐crossing location of the curve. A zero SHAP is an important indicator of a feature’s progressive importance for model prediction because a feature’s influence on prediction significantly shifts when its SHAP values go from positive to negative or vice versa. However, relying on zero‐crossing alone may be insufficient when the relationship is non‐linear. Consequently, we calculated the second derivatives to identify inflection points (peaks and troughs) along the curve. Lastly, we compiled all the potential cutpoints and performed post‐processing as shown in Fig. 1.ResultThe algorithm identified cutpoints at 1.44 SUVR for all and at 1.42 SUVR for CU participants (Fig. 2). At age 70, both cutpoints estimated the percentage of abnormals at less than 40% (Fig. 3).ConclusionTPE amyloid‐PET SUVR cutpoints for CUs equal the reported RW values (1.42 SUVR). A strength of the TPE is its application on cross‐sectional data making it versatile for various biomarkers across different cohorts.