Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Jun 1.
Published in final edited form as: Neurol Open Access. 2026 Apr 13;2(2):e000093.

Automated Imaging Differentiation for Dementia: Including Alzheimer Disease Dementia and Dementia with Lewy Bodies

Robin Chen 1, Shannon Y Chiu 2, Jesse C DeSimone 3, Wei-En Wang 3, Angelos Barmpoutis 4, Corey T Mcmillan 5, Hamsanandini Radhakrishnan 5, David John Irwin 5, Lyles Clark 5, Kejal Kantarci 6, Bradley F Boeve 6, David E Vaillancourt 3,7
PMCID: PMC13196839  NIHMSID: NIHMS2152182  PMID: 42179740

Abstract

BACKGROUND AND OBJECTIVES:

Differentiation of Alzheimer’s disease dementia (ADD) and dementia with Lewy bodies (DLB) remains a challenge. Free-water imaging has been investigated in neurodegenerative diseases and was found to be associated with neurodegeneration and neuroinflammation. This retrospective cohort study tested whether Automated Imaging Differentiation for Dementia (AIDD), combining diffusion free-water imaging (FWI) and support vector machine, predicts ADD vs DLB with high accuracy.

METHODS:

Diffusion MRI data was rendered from ADNI, NACC, and PDBP. Free-water and free-water corrected fractional anisotropy were calculated for each participant using a bi-tensor model. Diffusion metrics were randomly assigned to training and testing sets. The primary outcome was the area under the curve (AUC) in the test set. AIDD was paired with antemortem MRI to predict postmortem pathology.

RESULTS:

A total of 519 diffusion scans were processed with 258 ADD (mean age 73.7 (8.8), 50% male), 129 DLB (mean age 69.3, 88% male), and 132 controls (mean age 73.6 (6.8), 40% male). The machine learning sample included 387 scans,129 ADD with a mean age of 72.8 (8.7), 52.7% male; 129 DLB with a mean age of 69.3 (8.1), 87.6% male; and 129 controls with a mean age of 73.7 (6.8), 39.5% male). AIDD showed high training AUC for ADD vs DLB = 0.995 (95% CI, 0.985–1.000), ADD vs controls = 0.992 (95% CI, 0.982–1.000), DLB vs controls = 0.991 (95% CI, 0.983–0.999), and controls vs ADD/DLB = 0.990 (95% CI, 0.979–1.000). The testing AUCs were similar: ADD vs DLB = 0.995, ADD vs controls = 0.958, DLB vs controls = 0.939, controls vs ADD/DLB = 0.903. AIDD predictions were confirmed pathologically in a cohort of 13 patients.

DISCUSSION:

This study demonstrates that machine learning in combination with free-water imaging can differentiate ADD, DLB, and normal aging with high clinical and pathological accuracy. Advancement in early detection of dementia can lead to more appropriate treatment plans, especially for DLB, and improved disease stratification that have hindered drug development trials.

CLASSIFICATION of EVIDENCE:

This study provides Class II evidence that Automated Imaging Differentiation for Dementia, combining diffusion free-water imaging and machine learning accurately distinguishes Alzheimer’s Disease from Dementia with Lewy bodies.

Search Terms: [ 26 ] Alzheimer’s disease, [ 28 ] Dementia with Lewy bodies, [ 128 ] DWI, [ 322 ] Class II

INTRODUCTION:

Two most common causes of dementia in older adults are Alzheimer’s disease dementia (ADD) and dementia with Lewy bodies (DLB).1,2 Differentiating between ADD and DLB in the clinical environment remains challenging with high rates of misdiagnosis using the current standard of care.2 Up to 50% of neuropathologically-confirmed DLB, known as Lewy body disease (LBD), are correctly diagnosed antemortem, with ADD as the most common misdiagnosis.2,3 Distinguishing DLB from ADD is a vital part of patient care as DLB has a worse prognosis and requires different treatment plans compared to ADD.4 Patients with DLB are particularly sensitive to neuroleptics prescribed in dementia care, leading to worsening cognitive and motor functions.5 Further, new disease-modifying therapies are approved for ADD, but not for DLB.6,7

The National Institute on Aging and Alzheimer’s Association developed a research framework for Alzheimer’s disease (AD) classification using biomarkers such as amyloid, tau, and neurodegeneration.8 Amyloid positivity, as assessed using PET or biofluid assays (e.g., AB42/40, ptau217), is a core pathological, distinguishing feature of AD. However, amyloid and Lewy body co-pathologies occur in over 50% of LBD patients and can contribute to diagnostic uncertainty.2,9,10 In lieu of a DLB biomarker classification framework, current diagnostic criteria recommend combining indicative and supportive biomarkers to improve distinguishing between DLB and ADD. Indicative biomarkers include dopamine transporter scans (DaTscan), myocardial scintigraphy, and polysomnography. Supportive biomarkers are collected using MRI, PET, or SPECT scans, and EEG. Current MRI biomarkers in DLB leverages the relative sparing of the medial temporal lobe (MTL) to aid in differentiation.11,12 While biomarkers can aid differentiation, there is no singular biomarker that is specific to DLB.13 Thus, distinguishing between ADD and DLB relies on clinical follow-ups and multiple invasive biomarkers.

Recent data indicates that diffusion free-water imaging (FWI) is sensitive to disease-specific neurodegeneration in ADD vs DLB.14–17 Diffusion MRI is commonly obtained as part of the current standard of care but is not proactively used to help differentiate between ADD and DLB.18 In the Automated Imaging Differentiation for Parkinsonism (AIDP), a recently published 21-site prospective study showed that machine learning (ML) and FWI revealed >95% accuracy in distinguishing progressive supranuclear palsy and multiple system atrophy from Parkinson’s disease and differentiating multiple system atrophy from progressive supranuclear palsy.19 AIDP leveraged free-water (FW) and free-water corrected fractional anisotropy (FAT) values across regions in the brain to distinguish disease-specific diffusion pattern in Parkinson’s disease, multiple system atrophy, and progressive supranuclear palsy.19–23 At present, there is no singular antemortem biomarker that can definitively differentiate DLB from ADD.13 To expand on AIDP,19 we propose the Automated Imaging Differentiation for Dementia (AIDD). We propose using diffusion imaging with ML to assess different patterns of water movement in the brain between patients with ADD and patients with DLB as well as age-matched healthy controls. Here, we included healthy controls to assess the effects of normal aging on the brain in ML. In this study, we aim to explore the ability of FWI and ML, called AIDD, to distinguish DLB and ADD using clinically available pulse sequences and using multi-site imaging data obtained from three common scanner vendors.

METHODS:

Study Participants

This was a retrospective cohort study using research data from the Parkinson’s Disease Biomarker Program (PDBP), the National Alzheimer’s Coordinating Center (NACC), and the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Retrospective data were collected between January 2007 to March 2022. The ML development cohort included participants with a diagnosis of ADD, DLB,12 and controls without a history of neurological or neuropsychiatric conditions were eligible for study inclusion. Subjects with ADD met criteria based on Diagnostic and Statistical Manual of Mental Disorders (DSM-IV), 10th revision of the International Classification of Diseases (ICD-10), and/or the National Institute of Neurological and Communicative Disorders and Stroke – Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA. In ADNI, clinical consensus diagnoses incorporated these criteria along with cognitive measures (e.g., clinical dementia rating (CDR), Mini-Mental State Exam (MMSE), Montreal Cognitive Assessment (MoCA) with specific cutoffs for AD). In NACC, diagnoses were determined with a consensus panel or a single clinician at participating Alzheimer’s Disease Centers (ADCs), using patient history and standardized cognitive measures. All DLB cases met either the Third or Fourth Consortium criteria for probable DLB, and were determined by consensus panels within participating institutions.12 In addition, several pathologically-confirmed cases of AD and LBD from PDBP and NACC were included in this cohort and we will refer to the groups by clinical syndrome (ADD or DLB). The inclusion criteria included confirmation of amyloid PET positive status for the ADD cohort and confirmation of amyloid-PET negative status for the DLB and control cohorts. Pittsburgh compound B (PiB), 18F-Florbetapir, and 18F-Flortaucipir, and 18F-Florbetaben were used for amyloid PET for NACC and ADNI. PDBP solely used PiB. Standard uptake value ratio was used to determine amyloid status. General exclusion criteria included history of stroke, tumor, or other serious neurological conditions, and imaging data not compatible with the ML procedure described below. Incompatible imaging data included lower field strengths, missing b-value and b-vector pair, and scans with less than 6 b=1000 volumes. Diagnostic exclusions included amyloid-negative ADD, amyloid-positive DLB, and mild cognitive impairment (MCI) due to Lewy bodies. The combined cohort was examined for potential co-enrollment between sites, using a fuzzy lookup algorithm to check for high similarities between age, sex, ethnicity, and race. Baseline demographics such as sex, race (e.g., White), and ethnicity (i.e., Hispanic/non-Hispanic) were routinely obtained as part of the NACC Uniform Data Set (UDS). Demographics are self-reported but can be supplemented by proxy co-participant if needed. The pathological cohort included 5 patients with autopsy-confirmed AD without concomitant Lewy body neuropathology and 8 individuals with autopsy-confirmed LBD without concomitant AD neuropathology from the University of Pennsylvania brain bank. The pathological cohort was a different site and scanner from the clinical cohort, thus providing an independent test from the original scanners involved in training.

Longitudinal data for the DLB cohort from the PBDP consortium were incorporated into the analysis to account for a relatively small participant sample size compared to the ADD and control cohorts. For ADNI and NACC data, only one time point was included for data closest to the baseline visit. As described later, we also ran the analysis without any longitudinal data to determine if this was a factor.

Standard Protocol Approvals, Registrations, and Patient Consent

The study was approved by the Institutional Review Board at the University of Florida and all participants signed informed consent prior to data upload to PDBP, NACC, and ADNI. The study was performed in ethical accordance with the Declaration of Helsinki.

Diffusion Weighted Imaging

Diffusion MRI scans were acquired on Siemens (Erlangen, Germany), Philips (Best, Netherlands), and General Electric (Wisconsin, United States) 3T scanners using an echo planar imaging sequence with the following acquisition parameters: repetition time (TR) = 3230–16700 ms, echo time (TE) = 56–106 ms, slices = 81, 0 mm gap, flip angle = 90o, b0 images = 1–13, max b = 1000 s/mm2 (eTable 1). It is clear from eTable 1 that a wide range of scanners and scanner parameters were included. The distribution of scanners across each consortium is displayed on eTable 2. These imaging parameters are consistent with the ADNI-3 protocol and can be implemented on widely available 3T systems.24 Multi-shell images were converted to single-shell images by isolating the b = 1000 s/mm2 (b1000) volumes and discarding non-b1000 volumes.

Data processing was identical to previous works from our group.14,15,19,23. The diffusion processing pipeline included motion and eddy-current correction, skull-stripping, diffusion tensor fitting, free-water estimation, and normalization to Montreal Neurological Institute (MNI) space. The pipeline used custom UNIX shell scripts with FMRIB software library, Advanced Normalization Tools,25 and custom MATLAB (version R2020b) scripts 26,27 Quality control was done through automated signal-to-noise calculations and visual inspection of each diffusion map. Partial brains and imaging distortions were excluded from this analysis.

We used 20 ADD and 20 controls from ADNI to investigate the effects of susceptibility-induced geometric distortions and reported the intraclass correlation coefficients (ICCs) for key regions of interest.

The variables for ML included voxel-wise averages of FW and FAT from 120 regions of interest (ROIs). 108 ROIs were from the Mayo Clinic Adult Lifespan Template (MCALT) that included the forebrain, cerebellum, and hippocampus, and occipital, parietal, temporal, and entorhinal cortices. 28 Six custom ROIs were included and derived from previous work involving the midbrain.23 Six additional ROIs were included from prior work on tractography that encompassed transcallosal, cerebellothalamocortical, and locus-coeruleus-to-entorhinal-cortex tracts.17,29,30 All ROIs were normalized to MNI space and listed in eTable 3.

Machine Learning

We trained AIDD to perform disease-specific classification of four primary endpoints. Endpoints are comprised of three one-versus-one classifications (i.e., ADD vs DLB, ADD vs controls, DLB vs controls) and one one-versus-all classification (i.e., controls vs ADD/DLB). The one-versus-one method preserves the dataset balance but can be computationally intensive based on the number of comparisons. Conversely, one-versus-all trains on imbalanced datasets but is less computationally intensive.

FW and FAT metrics across 120 template regions and tracts were primary input variables for the training and validation of the model. Age and sex were not used as model features. The training and validation set included 80% of the total data, whereas the remaining 20% was held out for independent testing. Patients were randomly assigned to training and validation or test using stratified sampling to maintain sample size proportions. During random assignment, we constrained the algorithm to prevent any longitudinal datasets divided between train and test cohorts to prevent data leakage. In the training and validation, the data was randomly partitioned into five subgroups for five-fold cross-validation. Through each iteration, one of the five partitions is used as a validation set while the other four are used to train the model. In this process, the best hyperparameter (C parameter) for the support vector machine is identified and implemented to optimize the F-score, the harmonic mean of precision and recall. Then, we implemented the best hyperparameter in the final support vector model. A linear kernel support vector machine was selected for its interpretability and high performance. The trained model was evaluated on the testing dataset by comparing predictions to the clinical disease diagnoses using a 0.5 cutoff. For the neuropathology-confirmed cohort, we investigated the diffusion weighted imaging (DWI) scan closest to the autopsy from each patient to evaluate our AIDD model discriminative performance for ADD vs DLB.

In addition, we trained and evaluated SVMs that included age and sex with FW and FAT. The ML methodology is unaltered except for the data input.

To assess the influence of the inclusion of longitudinal DLB scans, we conducted a different analysis by retraining and retesting the ML models with only baseline scans and no longitudinal scans. All other aspects of the ML pipeline were identical.

To assess the potential influence of site effects, we conducted an alternative analysis. We hypothesize that if the SVM is learning site effects then the models should perform well with unrelated region of interests. Following a 3D t-test between controls, ADD and DLB, we identified 120 regions of interest not suspected to be different between the population. The 2 mm spherical ROIs are listed in eTable 4. New SVMs were developed on the new ROIs using the same data methodology and tested on the 20% holdout and neuropathology-confirmed cohort.

Statistical Analysis

For each endpoint, we conducted receiver operating characteristic (ROC) analysis to evaluate the discriminative performance of AIDD in training and validation and in testing. The area under the ROC curve (AUC) ranges from 0 to 1 and provides a description of the tradeoff in sensitivity and specificity at different cutoffs, with larger values indicating better classification. A bootstrapping approach with 1000 iterations was used to determine confidence intervals for AUCs. In addition, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated for each endpoint.

Data Availability

ADNI database and diagnostic information can be requested at ida.loni.usc.edu. NACC database and diagnostic details can be requested at docs.naccdata.org. PDBP database can be requested at pdbp.ninds.nih.gov.

RESULTS:

Demographic Data and Imaging Processing

A total of 730 diffusion MRI scans were identified for inclusion and 211 were excluded due to clinical and biomarker dissent for ADD or DLB, diagnosis of MCI due to Lewy bodies, or MRI acquisition parameters not compatible with the image processing and ML procedure (Figure 1). The final cohort included 258 ADD (age 73.7, 50% male), 129 DLB (age 69.3, 88% male), and 132 healthy controls (age 73.6, 60% female) (Table 1). 34 patients with DLB had more than one image. Regarding potential co-enrollment, while there were similarities on age and sex, ethnicity and race did not match, thus we found no matches across consortia indicating all patient samples were unique. For the PDBP cohort, the age was given as a range of a decade, thus, we approximated the age to be the median. This approximation was not done for the ADNI and NACC datasets. The distribution of patients is shown in Table 2.

Figure 1: Flow chart of exclusion criteria.

Figure 1:

Open in a new tab

This flow diagram shows the number of diffusion weighted images (DWI) accessed in this study. 211 DWIs were excluded from the study for not meeting criteria, resulting in 519 DWIs assessed in the study. Abbreviations: ADD = Alzheimer’s disease dementia, DLB = dementia with Lewy bodies, LB-MCI = mild cognitive impairment with Lewy bodies, and CON = healthy controls.

Table 1.

Patient demographics.

Total ADD DLB controls p-value
n (%) 519 258 (50) 129 (25) 132 (25)
Mean age (SD) - years 72.6 (8.4) 73.7 (8.8) 69.3 (8.1) 73.6 (6.8) <.0011
Sex - no. (%) <.0012
 Male 294 (57) 128 (50) 113 (88) 53 (40)
 Female 225 (43) 130 (50) 16 (12) 79 (60)
White 91.7% 91.1% 96.9% 87.9% 0.0232
Hispanic 9.1% 12.7% 1.5% 9.1% 0.0012
MoCA (SD) 18.8 (9.1) 15.5 (6.2) 20.0 (5.4) 26.0 (2.8) <.0011
Open in a new tab
1

ANOVA

2

Pearson’s Chi-square test

Abbreviations: ADD = Alzheimer’s disease dementia, DLB = dementia with Lewy bodies, MoCA = Montreal Cognitive Assessment, and SD=standard deviation.

Table 2:

Distribution of patients across different consortia.

Total ADNI NACC PDBP
n 519 214 182 123
Reference diagnosis - no. (%)
 Alzheimer’s disease dementia 258 (50) 82 (38) 175 (96) 1 (1)
 Dementia with Lewy bodies 129 (25) 0 7 (4) 122 (99)
 Controls 132 (25) 132 (62) 0 0
Open in a new tab

Abbreviations: ADNI = Alzheimer’s Disease Neuroimaging Initiative, NACC = National Alzheimer’s Coordinating Center, PDBP = Parkinson’s Disease Biomarker Program.

We reported high consistency, shown in ICC, between susceptibility-corrected scans versus scans without susceptibility correction. Additional information is on eTable 5,

Machine Learning Endpoints

ROC analyses on the training and validation cohorts and testing cohorts for AIDD are displayed in Figure 2. The ADD vs DLB model demonstrated high AUCs for training and validation (0.995, 95% confidence interval (CI) 0.985–1.000; sensitivity, 0.971; specificity, 1.000) and testing (0.995; sensitivity, 0.920; specificity, 0.960). Similarly, the ADD vs controls model had high AUCs for training and validation (0.992, 95% CI 0.982–1.000; sensitivity, 0.952; specificity, 0.990) and testing (0.958; sensitivity, 0.800; specificity, 0.920). DLB vs controls model had high AUCs in training and validation (0.991, 95% CI 0.983–0.999; sensitivity, 0.903; specificity, 0.971) and testing (0.939; sensitivity, 0.800; specificity, 0.960). The AUCs for controls vs ADD/DLB for training and validation (0.990, 95% CI 0.979–1.000; sensitivity, 0.923; specificity, 0.981) and testing (0.903; sensitivity, 0.840; specificity, 0.880) were high. Accuracy, sensitivity, specificity, positive predictive value, and negative predictive value are displayed in Table 3.

Figure 2: Receiver operator characteristics (ROC) curve of the four primary comparisons shown with training and validation set and testing set.

Figure 2:

Open in a new tab

The dash red line illustrates the performance of the model based on chance. Abbreviations: ADD = Alzheimer’s disease dementia, DLB = dementia with Lewy bodies, and CON = healthy controls.

Table 3.

Performance metrics for the four primary comparisons.

ADD v DLB ADD vs CON DLB vs CON CON vs ADD/DLB
Train (95% CI) Test Train (95% CI) Test Train (95% CI) Test Train (95% CI) Test
AUC 0.995 (0.985–1.000) 0.995 0.992 (0.982–1.000) 0.958 0.991 (0.983–0.999) 0.939 0.990 (0.979–1.000) 0.903
Accuracy 0.985 (0.966–1.000) 0.940 0.971 (0.947–0.990) 0.860 0.937 (0.903–0.966) 0.880 0.961 (0.939–0.980) 0.867
Sensitivity 0.971 (0.933–1.000) 0.920 0.952 (0.904–0.989) 0.800 0.903 (0.843–0.955) 0.800 0.923 (0.869–0.970) 0.840
Specificity 1.000 (1.000–1.000) 0.960 0.990 (0.968–1.000) 0.920 0.971 (0.937–1.000) 0.960 0.981 (0.960–0.995) 0.880
PPV 1.000 (1.000–1.000) 0.958 0.990 (0.967–1.000) 0.909 0.969 (0.935–1.000) 0.952 0.960 (0.916–0.990) 0.778
NPV 0.972 (0.935–1.000) 0.923 0.954 (0.909–0.990) 0.821 0.909 (0.847–0.956) 0.828 0.962 (0.936–0.986) 0.917
Open in a new tab

Abbreviations: ADD = Alzheimer’s disease dementia, DLB = dementia with Lewy bodies, CON = healthy controls, AUC = area under the curve, PPV = positive predictive value, NPV = negative predictive value.

The support vector machines trained with age, sex, FW, and FAT conveyed similar performance to the models trained on just FWI metrics. The ADD vs DLB model showed high AUCs for training and validation (0.993, 95% CI 0.981–1.000; sensitivity, 0.962; specificity, 1.000) and testing (0.931; sensitivity, 0.840; specificity, 0.960). The AUCs were high for the ADD vs controls model for training and validation (1.000, 95% CI 0.999–1.000; sensitivity, 0.981; specificity, 0.981) and testing (0.931; sensitivity, 0.840; specificity, 0.920). Similarly, the DLB vs controls model had high AUCs in training and validation (0.994, 95% CI 0.990–1.000; sensitivity, 0.970; specificity, 0.971) and testing (0.974; sensitivity, 0.800; specificity, 1.000). The AUCs were high for controls vs ADD/DLB for training and validation (0.993, 95% CI 0.988–1.000; sensitivity, 0.953; specificity, 0.971) and testing (0.910; sensitivity, 0.840; specificity, 0.800). Lastly, the ADD vs DLB model prediction on the neuropathology cohort predicted 12 out of 13 cases correctly (accuracy = 0.923).

The support vector machines without longitudinal data were trained on 204 individuals (68 individuals per diagnosis) and evaluated on 51 individuals (17 individuals per diagnosis). There were no longitudinal scans in this analysis. The ADD vs DLB model demonstrated high AUCs for training and validation (0.954, 95% CI 0.911–0.996; sensitivity, 0.955; specificity, 0.941) and testing (1.000; sensitivity, 0.941; specificity, 1.000). The ADD vs controls model had high AUCs for training and validation (0.989, 95% CI 0.974–1.000; sensitivity, 0.956; specificity, 0.956) and testing (0.993; sensitivity, 0.882; specificity, 1.000). The DLB vs controls model had high AUCs in training and validation (0.958, 95% CI 0.920–0.996; sensitivity, 0.853; specificity, 0.926) and testing (0.969; sensitivity, 0.941; specificity, 0.882). The controls vs ADD/DLB had high AUC for training and validation (0.984, 95% CI 0.972–0.997; sensitivity, 0.912; specificity, 0.934) and testing (0.959; sensitivity, 0.882; specificity, 0.941). The testing accuracy for ADD versus DLB is 97.1%, ADD versus controls is 94.1%, DLB versus controls is 91.2%, and controls versus ADD and DLB is 92.2%. The results reveal high prediction accuracy with or without longitudinal data.

SVMs trained on alternative ROIs showed high AUC for ADD vs DLB in training and validation set (0.987, 95% CI 0.972–1.000) but low AUC in the testing set (0.745). For ADD vs controls, 0.962, 95% CI 0.941–0.984 and 0.916. For DLB vs controls, 0.946, 95% CI 0.918 −0.973 and 0.769. For controls vs ADD/DLB, 0.852, 95% CI 0.808–0.896 and 0.916. The testing accuracies are 0.680, 0.796, 0.778, and 0.773 for ADD vs DLB, ADD vs controls, DLB vs controls, and controls vs ADD/DLB, respectively. Accuracy = 0.615, AUC = 0.925, sensitivity = 1.000, and specificity = 0.375 for the neuropathology cohort.

Regions of Interest Feature Importance

We extracted feature weights for each comparison. We examined the top five regions for the positive class and the top five regions for the negative class (e.g., ADD (positive) vs DLB (negative)). The analysis of the regions was projected onto a glass brain in MNI space, displayed in Figure 3. The top regions for ADD vs DLB are right cerebellum lobule VIII, left cerebellum lobule IX, right entorhinal cortex, right middle orbitofrontal gyrus, bilateral olfactory cortices, pons, left supplementary motor area, left locus-coeruleus-to-transentorhinal-cortex tract, and splenium of the corpus callosum. The top regions for ADD vs controls are the right amygdala, right angular gyrus, left calcarine gyrus, left cerebellar lobules VIIb and VIII, bilateral dorsal dentate, left Heschl’s gyrus, left and right hippocampus, and right superior occipital gyrus. The top regions for DLB vs controls are the dorsal mesopontine tegmentum, left Heschl’s gyrus, left and right pallidum, posterior substantia nigra, right inferior temporal gyrus, right superior temporal pole, vermis lobules VIII and IX, and splenium of the corpus callosum. The top regions for controls vs ADD/DLB are left cerebellum Crus II, left insula, right medial frontal gyrus, left middle frontal gyrus, left Heschl’s gyrus, right parahippocampal gyrus, left putamen, middle temporal lobe, and genu of the corpus callosum. While we only describe the top ten regions in this study, the SVM predictions were not solely derived from these regions but rather from a network of FW and FAT values across the brain. The feature weights do not implicitly imply a relationship between the disease group though it does indicate the relative contribution associated with an elevated FW or FAT.

Figure 3: Glass brain representation of the top five regions of interest (ROIs) indicated for the primary comparisons.

Figure 3:

Open in a new tab

The first column shows the anterior-left view, and the second column shows the posterior-right view. The dark red regions indicate higher FW, and the orange region indicates higher FAT for the positive class. The dark blue regions indicate lower FW, and the light blue region indicates lower FAT for the negative class. There was no overlap between FW and FAT regions. (A) ADD vs DLB, (B) ADD vs controls, (C) DLB vs controls, and (D) controls vs ADD/DLB. Abbreviations: ADD = Alzheimer’s disease dementia, DLB = dementia with Lewy bodies, FW = free-water, FAT = free-water-corrected fractional anisotropy.

Neuropathology Predictions

The median time between the last clinical imaging scan and autopsy in the pathology cohort was 4 years (range, 0–9 years). The neuropathology cohort consisted of 8 LBDs and 5 ADs (Table 4). AIDD predicted the autopsy-confirmed neuropathological diagnosis in 13 of 13 brains (100%), including 8 of 8 LBD brains and 5 of 5 AD brains. AIDD metrics demonstrated an overall diagnostic accuracy gain of 46.2% compared to the last available antemortem clinical diagnosis, which had an accuracy of only 53.8% (6 of 8 LBD and 1 of 5 AD).

Table 4:

Application of the AIDD model for predicting pathological diagnosis in 13 patients.

CERAD Thal stage Braak stage LB Pathology Clinical Path. ADD vs DLB (ADD probability)
1 0 0 1 Diffuse or Neocortical DLB LBD DLB (0.037)
2 0 0 1 Diffuse or Neocortical DLB LBD DLB (0.194)
3 0 0 2 Brainstem Predominant PD-MCI LBD DLB (0.102)
4 0 0 4 Brainstem Predominant PDD LBD DLB (0.073)
5 0 0 2 Diffuse or Neocortical CBS LBD DLB (0.359)
6 0 0 2 Transitional or Limbic PDD LBD DLB (0.201)
7 0 0 2 Diffuse or Neocortical DLB LBD DLB (0.204)
8 0 0 1 Brainstem Predominant CBS LBD DLB (0.031)
9 3 3 6 None CBS AD ADD (0.993)
10 3 3 6 None ADD AD ADD (0.991)
11 2 2 4 None CBS AD ADD (0.506)
12 3 3 6 Amygdala Predominant CBS AD ADD (0.914)
13 3 3 6 None CBS AD ADD (0.625)
Open in a new tab

Abbreviations: ADD = Alzheimer’s disease dementia, CERAD = consortium to establish a registry for Alzheimer’s disease, DLB = dementia with Lewy bodies, MCI = mild cognitive impairment, path. = pathological diagnosis

Classification of Evidence

This study provides Class II evidence that Automated Imaging Differentiation for Dementia, combining diffusion free-water imaging and machine learning accurately distinguishes Alzheimer’s Disease from Dementia with Lewy bodies.

DISCUSSION:

This retrospective cohort study of AIDD demonstrated excellent disease-specific classification of amyloid-positive ADD and amyloid-negative DLB. AIDD demonstrated excellent testing AUCs across four primary endpoints. Strengths of this study included the use of clinical and biomarker information to establish a reliable ground truth diagnosis for ML, the use of a large retrospective cohort from three separate consortia, the use of a single, clinically viable diffusion MRI scan collected on three different and common scanner vendors, and independent validation of AIDD in a brain bank cohort with neuropathological confirmation of AD and LBD.

Current FDA-approved disease-modifying therapies specifically target amyloid which is present in all patients with ADD and in half of all patients with DLB.10 While anti-amyloid therapies may have benefits for amyloid-positive DLB patients, they are not approved for use outside of AD-specific amyloidosis. Amyloid status can be determined using blood assays, cerebrospinal fluid assays, or amyloid PET. However, these tests are not specific to ADD. Another test, DaTscan, could distinguish DLB from ADD, however, its sensitivity is low (77.7%).31 While amyloid PET/SPECT and DaTscans are used, Armstrong and others observed that DLB specialists most commonly order MRI despite that MRI cannot confirm a DLB diagnosis.32 Given the availability of MRIs, AIDD could serve as an alternative to multiple tests, specifically the radiologic imaging tests, to be used in conjunction with neuropsychological evaluations. Rather than employ multiple tests to evaluate beta-amyloid levels, AIDD uses a single MRI acquisition to distinguish ADD and DLB.

We report on our investigation on FWI and ML between ADD and DLB. Studies have used T1-weighted MRIs to study the differences in structural features, such as cortical thickness, between ADD and DLB.11,33 Vemuri and others used ROI analyses with T1-weighted MRI and showed a differential accuracy of 87% for AD, 95% for LBD and 90% for frontotemporal lobar degeneration.34 Wang et al. combined deep learning with T1-weighted measures and showed a balanced accuracy of 0.844 for AD and 0.623 for LBD using a cohort from NACC and ADNI.35 Further, Lebedev et al. used cortical thickness to differentiate between AD and DLB and showed AUCs of 0.731 for the combined testing datasets and 0.670 and 0.560 when training on one cohort and testing on the other as an independent dataset.36 MTL atrophy has been identified to be greater in ADD than in both DLB and controls. Matsuda et al. concurs, showing more degeneration in the medial temporal lobe in ADD compared to DLB using voxel-based morphometry, however the discrimination between AD and DLB was 63.3% on the test set of 210 patients.37 Feature weight analysis showed that elevated FW in the right and left MTLs are more indicative of ADD compared to DLB, although, FW elevation in the olfactory regions demonstrated the greatest level of feature importance for the ADD vs DLB endpoint. Multiple studies have reported sparing of the hippocampus in DLB compared to ADD.11,38 Fractional anisotropy (FA) in the pons and left thalamus was reported to be lower in DLB vs ADD using conventional diffusion techniques.11 We observed lower FAT in the pons but not in the left thalamus. Further, we observed that elevated FW values in the olfactory regions suggest DLB diagnosis. The degeneration in the olfactory cortices relates to symptoms of hyposmia and anosmia commonly associated with Lewy body pathology.39–42 Overall, our results agree with the reported differences between ADD and DLB.

In the ADD vs controls model, the features with the greatest relative importance included free-water in the hippocampus, amygdala, angular gyrus, dorsal dentate, calcarine sulcus, and cerebellar lobules VIIb and VII and FAT in the superior occipital lobe and Heschl’s gyrus. These results agree with published literature investigating AD pathological changes. Changes in the limbic regions, particularly the hippocampus, correlated with AD pathology.43 Our results concur by showing elevated FW in the hippocampus is more indicative of ADD than controls. Other regions outside of the top five regions included the amygdala and segments of the corpus callosum. We found elevated free-water in the right amygdala contributed to ADD probability. This finding is consistent with literature on the neuropathological limbic changes.15,44 FA and FAT are lower in the corpus callosum, except the splenium, when comparing ADD to controls.45 We observed lower FAT in the genu and the splenium, but not in the body. 45

In DLB vs controls, the top features were FW in the dorsal mesopontine tegmentum and Heschl’s gyrus, and FAT in the globus pallidus, posterior substantia nigra, inferior temporal lobe, vermis lobule VIII, superior temporal pole, and the splenium of the corpus callosum. Many studies have reported on the dopaminergic pathway difference between DLB and controls, which we observe in our midbrain ROIs. Additional regions included the precuneus and cingulum. We saw lower FAT in the right precuneus and the cingulum was more indicative of DLB than controls. In addition, we see elevated free-water in the left and right precuneus in DLB vs controls. This result is consistent with published studies, including reduced FA in the precuneus and cingulate cortex,46 and reduced volume in the cingulum.47

When combining the ADD and DLB group, the controls vs ADD and DLB model showed elevated free-water in the insula, Heschl’s gyrus, genu of the corpus callosum, and decreased free-water in the parahippocampal gyrus, and cerebellum Crus II, and increased FAT in the left middle frontal gyri and putamen, and decreased FAT in the medial frontal gyri, and left middle temporal gyrus. MTL is a key point of interest, noted by other studies, to show more atrophy in ADD and DLB compared to controls, with more preservation in DLB compared to ADD. The controls vs ADD and DLB model had a lower performance compared to the one-versus-one models (ADD vs controls and DLB vs controls). This is likely attributed to the data imbalance in the training as we combined the ADD and DLB groups. Data imbalance can lead to poor performance for the minority class, controls in this case.

Overall, the regions agree with pathological reports on DLB and ADD. Further, we tested the ADD vs DLB model on thirteen patients with autopsy-confirmed diagnoses. 80% of the subjects with AD pathology were diagnosed as corticobasal syndrome (CBS) clinically whereas 25% of the LBD were diagnosed as CBS. The neuropathology underlying CBS is widely heterogeneous, with corticobasal degeneration (CBD) as the most common, therefore necessitating the use of various biomarkers to differentiate.48,49 AIDD performed with 100% accuracy in a small pathological cohort, which suggests good utility for DLB without concomitant AD pathology.

There are limitations that deserve discussion. First, our defined separation between ADD and DLB was limited to amyloid positivity, and we did not assess the synuclein status for the clinical cohort. Future studies could investigate the symptoms of dementia in relation to the AIDD probability in cases with amyloid and synuclein co-pathologies. Second, blood-based biomarkers have shown substantial promise in relation to amyloid status and cognitive decline. While our current study did not include these biomarkers, future evaluations should combine AB42/40 and ptau217 values with AIDD. Third, the ethnic and racial compositions for this cohort were primarily non-Hispanic and white and the DLB cohort was disproportionately male. Increasing patient diversity can improve the generalizability of this study. Further, the SVM models were developed on 3T DWI scans, and cannot be generalized to 1.5 T and 7 T MRI scans. Lastly, future work will include mixed pathology cases, such as amyloid-positive DLB, which is a portion of the clinical population of DLB.50 As our study excludes amyloid-positive DLB cases, current result and performance may be inflated and not fully applicable to the spectrum of DLB cases with co-pathologies. Future work will incorporate more external datasets for further validation.

This study provides proof-of-concept evidence that machine learning with a single, non-invasive, non-ionizing, and clinically viable diffusion MRI scan can differentiate patients with ADD and DLB with high accuracy. AIDD showed a diagnostic gain of 46.2% in an autopsy-confirmed cohort, with improvements in both AD and LBD. The regions with larger feature importance such as the entorhinal cortex, hippocampus, and MTL, are supported in literature investigating imaging differences in ADD, DLB, and controls.

Supplementary Material

eTable 1
eTable 2
eTable 3
eTable 4
eTable 5

Acknowledgment:

Data and biospecimens used in preparation of this manuscript were obtained from the Parkinson’s Disease Biomarkers Program (PDBP) Consortium, supported by the National Institute of Neurological Disorders and Stroke at the National Institutes of Health. Investigators include: Roger Albin, Roy Alcalay, Alberto Ascherio, Thomas Beach, Sarah Berman, Bradley Boeve, F. DuBois Bowman, Shu Chen, Alice Chen-Plotkin, William Dauer, Ted Dawson, Paula Desplats, Richard Dewey, Ray Dorsey, Jori Fleisher, Kirk Frey, Douglas Galasko, James Galvin, Dwight German, Steven Gunzler, Lawrence Honig, Xuemei Huang, David Irwin, Kejal Kantarci, Anumantha Kanthasamy, Daniel Kaufer, Qingzhong Kong, James Leverenz, Allan Levey, Carol Lippa, Irene Litvan, Oscar Lopez, Jian Ma, Richard Mailman, Lara Mangravite, Karen Marder, Nandakumar Narayanan, Laurie Orzelius, Vladislav Petyuk, Judith Potashkin, Liana Rosenthal, Rachel Saunders-Pullman, Clemens Scherzer, Michael Schwarzschild, Nicholas Seyfried, Tanya Simuni, Andrew Singleton, David Standaert, Debby Tsuang, David Vaillancourt, Jerrold Vitek, David Walt, Andrew West, Cyrus Zabetian, and Jing Zhang. The PDBP Investigators have not participated in reviewing the data analysis or content of the manuscript. Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12–2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California. The NACC database is funded by NIA/NIH Grant U24 AG072122. SCAN is a multi-institutional project that was funded as a U24 grant (AG067418) by the National Institute on Aging in May 2020. Data collected by SCAN and shared by NACC are contributed by the NIA-funded ADRCs as follows: Arizona Alzheimer’s Center - P30 AG072980 (PI: Eric Reiman, MD); R01 AG069453 (PI: Eric Reiman (contact), MD); P30 AG019610 (PI: Eric Reiman, MD); and the State of Arizona which provided additional funding supporting our center; Boston University - P30 AG013846 (PI Neil Kowall MD); Cleveland ADRC - P30 AG062428 (James Leverenz, MD); Cleveland Clinic, Las Vegas - P20AG068053; Columbia - P50 AG008702 (PI Scott Small MD); Duke/UNC ADRC - P30 AG072958; Emory University - P30AG066511 (PI Levey Allan, MD, PhD); Indiana University - R01 AG19771 (PI Andrew Saykin, PsyD); P30 AG10133 (PI Andrew Saykin, PsyD); P30 AG072976 (PI Andrew Saykin, PsyD); R01 AG061788 (PI Shannon Risacher, PhD); R01 AG053993 (PI Yu-Chien Wu, MD, PhD); U01 AG057195 (PI Liana Apostolova, MD); U19 AG063911 (PI Bradley Boeve, MD); and the Indiana University Department of Radiology and Imaging Sciences; Johns Hopkins - P30 AG066507 (PI Marilyn Albert, Phd.); Mayo Clinic - P30 AG062677 (PI Ronald Petersen MD PhD); Mount Sinai - P30 AG066514 (PI Mary Sano, PhD); R01 AG054110 (PI Trey Hedden, PhD); R01 AG053509 (PI Trey Hedden, PhD); New York University - P30AG066512–01S2 (PI Thomas Wisniewski, MD); R01AG056031 (PI Ricardo Osorio, MD); R01AG056531 (PIs Ricardo Osorio, MD; Girardin Jean-Louis, PhD); Northwestern University - P30 AG013854 (PI Robert Vassar PhD); R01 AG045571 (PI Emily Rogalski, PhD); R56 AG045571, (PI Emily Rogalski, PhD); R01 AG067781, (PI Emily Rogalski, PhD); U19 AG073153, (PI Emily Rogalski, PhD); R01 DC008552, (M.-Marsel Mesulam, MD); R01 AG077444, (PIs M.-Marsel Mesulam, MD, Emily Rogalski, PhD); R01 NS075075 (PI Emily Rogalski, PhD); R01 AG056258 (PI Emily Rogalski, PhD); Oregon Health and Science University - P30 AG008017 (PI Jeffrey Kaye MD); R56 AG074321 (PI Jeffrey Kaye, MD); Rush University - P30 AG010161 (PI David Bennett MD); Stanford - P30AG066515; P50 AG047366 (PI Victor Henderson MD MS); University of Alabama, Birmingham - P20; University of California, Davis - P30 AG10129 (PI Charles DeCarli, MD); P30 AG072972 (PI Charles DeCarli, MD); University of California, Irvine - P50 AG016573 (PI Frank LaFerla PhD); University of California, San Diego - P30AG062429 (PI James Brewer, MD, PhD); University of California, San Francisco - P30 AG062422 (Rabinovici, Gil D., MD); University of Kansas - P30 AG035982 (Russell Swerdlow, MD); University of Kentucky - P30 AG028283–15S1 (PIs Linda Van Eldik, PhD and Brian Gold, PhD); University of Michigan ADRC - P30AG053760 (PI Henry Paulson, MD, PhD) P30AG072931 (PI Henry Paulson, MD, PhD) Cure Alzheimer’s Fund 200775 - (PI Henry Paulson, MD, PhD) U19 NS120384 (PI Charles DeCarli, MD, University of Michigan Site PI Henry Paulson, MD, PhD) R01 AG068338 (MPI Bruno Giordani, PhD, Carol Persad, PhD, Yi Murphey, PhD) S10OD026738–01 (PI Douglas Noll, PhD) R01 AG058724 (PI Benjamin Hampstead, PhD) R35 AG072262 (PI Benjamin Hampstead, PhD) W81XWH2110743 (PI Benjamin Hampstead, PhD) R01 AG073235 (PI Nancy Chiaravalloti, University of Michigan Site PI Benjamin Hampstead, PhD) 1I01RX001534 (PI Benjamin Hampstead, PhD) IRX001381 (PI Benjamin Hampstead, PhD); University of New Mexico - P20 AG068077 (Gary Rosenberg, MD); University of Pennsylvania - State of PA project 2019NF4100087335 (PI David Wolk, MD); Rooney Family Research Fund (PI David Wolk, MD); R01 AG055005 (PI David Wolk, MD); University of Pittsburgh - P50 AG005133 (PI Oscar Lopez MD); University of Southern California - P50 AG005142 (PI Helena Chui MD); University of Washington - P50 AG005136 (PI Thomas Grabowski MD); University of Wisconsin - P50 AG033514 (PI Sanjay Asthana MD FRCP); Vanderbilt University - P20 AG068082; Wake Forest - P30AG072947 (PI Suzanne Craft, PhD); Washington University, St. Louis - P01 AG03991 (PI John Morris MD); P01 AG026276 (PI John Morris MD); P20 MH071616 (PI Dan Marcus); P30 AG066444 (PI John Morris MD); P30 NS098577 (PI Dan Marcus); R01 AG021910 (PI Randy Buckner); R01 AG043434 (PI Catherine Roe); R01 EB009352 (PI Dan Marcus); UL1 TR000448 (PI Brad Evanoff); U24 RR021382 (PI Bruce Rosen); Avid Radiopharmaceuticals / Eli Lilly; Yale - P50 AG047270 (PI Stephen Strittmatter MD PhD); R01AG052560 (MPI: Christopher van Dyck, MD; Richard Carson, PhD); R01AG062276 (PI: Christopher van Dyck, MD); 1Florida - P30AG066506–03 (PI Glenn Smith, PhD); P50 AG047266 (PI Todd Golde MD PhD)

Study Funding:

This work was supported by the National Institutes of Health [grant numbers: U01NS100620, T32NS082168, P30AG066506].

Footnotes

Disclosure:

R. Chen reported no disclosures; S.Y. Chiu reported receiving grants from the NIH; J.C. DeSimone reported no disclosures; W. Wang reported no disclosures; A. Barmpoutis reported receiving grants from the NIH, and National Science Foundation, and being cofounder and shareholder of Neuropacs Corp outside the submitted work and having a patent application (WO/2025/071885); C.T. McMillan reported no disclosures; H. Radhakirshnan reported no disclosure; D.J. Irwin reported receiving grants from the NIH, Denali Therapeutics, Alector, Passage Bio, Prevail, and Cervomed, serving as unpaid advisor for the Lewy Body Disease Association (LBDA), and the Association for Frontotemporal Degeneration; L. Clark reported no disclosures; K. Kantarci reported receiving grants from the NIH, consultant fees from Biogen, Eisai, and BioArtic, and a gift from Eli Lilly; B. Boeve reported receiving grants from the NIH, LBDA, Mayo Clinic, American Brain Foundation, Little Family Foundation, Ted Turn and Family Foundation, Alector, EIP Pharma/Cervomed, Cognition Therapeutics, and Transposon, and advisor fees from the Tau Consortium; D.E. Vaillancourt reported receiving nonfinancial support from Automated Imaging Diagnostics, grants from the NIH, nonfinancial support from Department of Defense, cofounder and shareholder of Neuropacs Corp, and licensing a patent (11439341).

REFERENCES:

  • 1.Galvin JE. Lewy Body Dementia. CONTINUUM Lifelong Learning in Neurology. 2024;30(6):1673–1698. doi: 10.1212/CON.0000000000001496 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Sim J, Li H, Hameed S, Ting SKS. Clinical Manifestations of Early-Onset Dementia with Lewy Bodies Compared with Late-Onset Dementia with Lewy Bodies and Early-Onset Alzheimer Disease. JAMA Neurol. 2022;79(7):702–709. doi: 10.1001/jamaneurol.2022.1133 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Serrano GE, Shprecher D, Callan M, et al. Cardiac sympathetic denervation and synucleinopathy in Alzheimer’s disease with brain Lewy body disease. Brain Commun. 2020;2(1). doi: 10.1093/braincomms/fcaa004 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Mueller C, Soysal P, Rongve A, et al. Survival time and differences between dementia with Lewy bodies and Alzheimer’s disease following diagnosis: A meta-analysis of longitudinal studies. Ageing Res Rev. 2019;50(November 2018):72–80. doi: 10.1016/j.arr.2019.01.005 [DOI] [PubMed] [Google Scholar]
  • 5.Baskys A Lewy body dementia: The Litmus test for neuroleptic sensitivity and extrapyramidal symptoms. Journal of Clinical Psychiatry. 2004;65(SUPPL. 11):16–22. https://www.psychiatrist.com/read-pdf/10327/ [PubMed] [Google Scholar]
  • 6.Sims JR, Zimmer JA, Evans CD, et al. Donanemab in Early Symptomatic Alzheimer Disease: The TRAILBLAZER-ALZ 2 Randomized Clinical Trial. JAMA. 2023;330(6):512–527. doi: 10.1001/jama.2023.13239 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.van Dyck CH, Swanson CJ, Aisen P, et al. Lecanemab in Early Alzheimer’s Disease. New England Journal of Medicine. 2023;388(1):9–21. doi: 10.1056/NEJMoa2212948 [DOI] [PubMed] [Google Scholar]
  • 8.Jack CR, Bennett DA, Blennow K, et al. NIA-AA Research Framework: Toward a biological definition of Alzheimer’s disease. Alzheimer’s and Dementia.Elsevier Inc. 2018;14(4):535–562. doi: 10.1016/j.jalz.2018.02.018 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Ferreira D, Przybelski SA, Lesnick TG, et al. β-Amyloid and tau biomarkers and clinical phenotype in dementia with Lewy bodies. Neurology. 2020;95(24):e3257–e3268. doi: 10.1212/WNL.0000000000010943 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Irwin DJ, Grossman M, Weintraub D, et al. Neuropathological and genetic correlates of survival and dementia onset in synucleinopathies: a retrospective analysis. Lancet Neurol. 2017;16(1):55–65. doi: 10.1016/S1474-4422(16)30291-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Mak E, Su L, Williams GB, O’Brien JT, O’Brien JT. Neuroimaging characteristics of dementia with Lewy bodies. Alzheimers Res Ther. 2014;6(2):18. doi: 10.1186/alzrt248 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.McKeith IG, Boeve BF, Dickson DW, et al. Diagnosis and management of dementia with Lewy bodies: Fourth consensus report of the DLB Consortium. Neurology. 2017;89(1):88–100. doi: 10.1212/WNL.0000000000004058 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Abdelmoaty MM, Lu E, Kadry R, et al. Clinical biomarkers for Lewy body diseases. Cell Biosci. 2023;13(1):1–23. doi: 10.1186/s13578-023-01152-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Chiu SY, Chen R, Wang W, et al. Longitudinal Free‐Water Changes in Dementia with Lewy Bodies. Movement Disorders. 2024;39(5):836–846. doi: 10.1002/mds.29763 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Ofori E, DeKosky ST, Febo M, et al. Free-water imaging of the hippocampus is a sensitive marker of Alzheimer’s disease. Neuroimage Clin. 2019;24(June):101985. doi: 10.1016/j.nicl.2019.101985 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.DeSimone JC, Wang W en, Loewenstein DA, et al. Diffusion MRI relates to plasma Aβ42/40 in PET negative participants without dementia. Alzheimer’s and Dementia. 2024;20(4):2830–2842. doi: 10.1002/alz.13693 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Chu WT, Wang W en E, Zaborszky L, et al. Association of Cognitive Impairment With Free Water in the Nucleus Basalis of Meynert and Locus Coeruleus to Transentorhinal Cortex Tract. Neurology. 2021;98(7):10.1212/WNL.0000000000013206. doi: 10.1212/wnl.0000000000013206 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Benjamin P, Khan F, Mackinnon AD. The use of diffusion weighted imaging to evaluate pathology outside the brain parenchyma in neuroimaging studies. British Journal of Radiology. 2017;90(1072). doi: 10.1259/bjr.20160821 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Vaillancourt DE, Barmpoutis A, Wu SS, et al. Automated Imaging Differentiation for Parkinsonism. JAMA Neurol. 2025;82(5):495. doi: 10.1001/jamaneurol.2025.0112 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Pasternak O, Kubicki M, Shenton ME. In vivo imaging of neuroinflammation in schizophrenia. Schizophr Res. 2016;173(3):200–212. doi: 10.1016/j.schres.2015.05.034 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Pasternak O, Sochen N, Gur Y, Intrator N, Assaf Y. Free water elimination and mapping from diffusion MRI. Magn Reson Med. 2009;62(3):717–730. doi: 10.1002/mrm.22055 [DOI] [PubMed] [Google Scholar]
  • 22.Pasternak O, Westin CF, Bouix S, et al. Excessive extracellular volume reveals a neurodegenerative pattern in schizophrenia onset. Journal of Neuroscience. 2012;32(48):17365–17372. doi: 10.1523/JNEUROSCI.2904-12.2012 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Archer DB, Bricker JT, Chu WT, et al. Development and validation of the automated imaging differentiation in parkinsonism (AID-P): a multicentre machine learning study. Lancet Digit Health. 2019;1(5):e222–e231. doi: 10.1016/S2589-7500(19)30105-0 [DOI] [PubMed] [Google Scholar]
  • 24.Zavaliangos-Petropulu A, Nir TM, Thomopoulos SI, et al. Diffusion MRI indices and their relation to cognitive impairment in brain aging: The updated multi-protocol approach in ADNI3. Front Neuroinform. 2019;13(February):1–18. doi: 10.3389/fninf.2019.00002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Jenkinson M, Beckmann CF, Behrens TEJ, Woolrich MW, Smith SM. FSL. Neuroimage. 2012;62(2):782–790. doi: 10.1016/j.neuroimage.2011.09.015 [DOI] [PubMed] [Google Scholar]
  • 26.Pasternak O, Shenton ME, Westin CF. Estimation of extracellular volume from regularized multi-shell diffusion MRI. In: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol 7511 LNCS. 2012:305–312. doi: 10.1007/978-3-642-33418-4_38 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Metzler-Baddeley C, O’Sullivan MJ, Bells S, Pasternak O, Jones DK. How and how not to correct for CSF-contamination in diffusion MRI. Neuroimage. 2012;59(2):1394–1403. doi: 10.1016/j.neuroimage.2011.08.043 [DOI] [PubMed] [Google Scholar]
  • 28.Schwarz CG, Gunter JL, Ward CP, et al. The Mayo Clinic Adult Lifespan Template: Better Quantification Across the Lifespan. Alzheimer’s & Dementia. 2017;13(7S_Part_16):P792. doi: 10.1016/j.jalz.2017.06.1071 [DOI] [Google Scholar]
  • 29.Archer DB, Coombes SA, McFarland NR, DeKosky ST, Vaillancourt DE. Development of a transcallosal tractography template and its application to dementia. Neuroimage. 2019;200(March):302–312. doi: 10.1016/j.neuroimage.2019.06.065 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Archer DB, Coombes SA, Chu WT, et al. A widespread visually-sensitive functional network relates to symptoms in essential tremor. Brain. 2018;141(2):472–485. doi: 10.1093/brain/awx338 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Ye BS, Lee S, Yoo H, et al. Distinguishing between dementia with Lewy bodies and Alzheimer’s disease using metabolic patterns. Neurobiol Aging. 2020;87:11–17. doi: 10.1016/j.neurobiolaging.2019.10.020 [DOI] [PubMed] [Google Scholar]
  • 32.Armstrong MJ, Irwin DJ, Leverenz JB, Gamez N, Taylor A, Galvin JE. Biomarker use for dementia with lewy body diagnosis: Survey of US experts. Alzheimer Dis Assoc Disord. 2021;35(1):55–61. doi: 10.1097/WAD.0000000000000414 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Wang WY, Yu JT, Liu Y, et al. Voxel-based meta-analysis of grey matter changes in Alzheimer’s disease. Transl Neurodegener. 2015;4(1):1–9. doi: 10.1186/s40035-015-0027-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Vemuri P, Simon G, Kantarci K, et al. Antemortem differential diagnosis of dementia pathology using structural MRI: Differential-STAND. Neuroimage. 2011;55(2):522–531. doi: 10.1016/j.neuroimage.2010.12.073 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Wang D, Honnorat N, Toledo JB, et al. Deep learning reveals pathology-confirmed neuroimaging signatures in Alzheimer’s, vascular and Lewy body dementias. Brain. Published online December 9, 2024:1–15. doi: 10.1093/brain/awae388 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Lebedev A V, Westman E, Beyer MK, et al. Multivariate classification of patients with Alzheimer’s and dementia with Lewy bodies using high-dimensional cortical thickness measurements: An MRI surface-based morphometric study. J Neurol. 2013;260(4):1104–1115. doi: 10.1007/s00415-012-6768-z [DOI] [PubMed] [Google Scholar]
  • 37.Matsuda H, Yokoyama K, Sato N, et al. Differentiation between dementia with lewy bodies and alzheimer’s disease using voxel-based morphometry of structural MRI: A multicenter study. Neuropsychiatr Dis Treat. 2019;15:2715–2722. doi: 10.2147/NDT.S222966 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38.Murray ME, Ferman TJ, Boeve BF, et al. MRI and pathology of REM sleep behavior disorder in dementia with Lewy bodies. Neurology. 2013;81(19):1681–1689. doi: 10.1212/01.wnl.0000435299.57153.f0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.McShane RH, Nagy Z, Esiri MM, et al. Anosmia in dementia is associated with Lewy bodies rather than Alzheimer’s pathology. J Neurol Neurosurg Psychiatry. 2001;70(6):739–743. doi: 10.1136/jnnp.70.6.739 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Williams SS, Williams J, Combrinck M, Christie S, Smith AD, McShane R. Olfactory impairment is more marked in patients with mild dementia with Lewy bodies than those with mild Alzheimer disease. J Neurol Neurosurg Psychiatry. 2009;80(6):667–670. doi: 10.1136/jnnp.2008.155895 [DOI] [PubMed] [Google Scholar]
  • 41.Beach TG, White CL, Hladik CL, et al. Olfactory bulb α-synucleinopathy has high specificity and sensitivity for Lewy body disorders. Acta Neuropathol. 2009;117(2):169. doi: 10.1007/s00401-008-0450-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42.Coughlin DG, Macleod KR, Middleton JS, et al. Association of CSF α-Synuclein Seeding Amplification Assay Results with Clinical Features of Possible and Probable Dementia with Lewy Bodies. Neurology. 2024;103(3):1–11. doi: 10.1212/WNL.0000000000209656 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43.Braak H, Braak E. Neuropathological stageing of Alzheimer-related changes. Acta Neuropathol. 1991;82(4):239–259. doi: 10.1007/BF00308809 [DOI] [PubMed] [Google Scholar]
  • 44.Yamashiro K, Takabayashi K, Kamagata K, et al. Free water in gray matter linked to gut microbiota changes with decreased butyrate producers in Alzheimer’s disease and mild cognitive impairment. Neurobiol Dis. 2024;193(March):106464. doi: 10.1016/j.nbd.2024.106464 [DOI] [PubMed] [Google Scholar]
  • 45.Bergamino M, Walsh RR, Stokes AM. Free-water diffusion tensor imaging improves the accuracy and sensitivity of white matter analysis in Alzheimer’s disease. Sci Rep. 2021;11(1):1–12. doi: 10.1038/s41598-021-86505-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46.Watson R, Blamire AM, Colloby SJ, et al. Characterizing dementia with Lewy bodies by means of diffusion tensor imaging. Neurology. 2012;79(9):906–914. doi: 10.1212/WNL.0b013e318266fc51 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47.Sanchez-Castaneda C, Rene R, Ramirez-Ruiz B, et al. Correlations between gray matter reductions and cognitive deficits in dementia with lewy bodies and Parkinson’s disease with dementia. Movement Disorders. 2009;24(12):1740–1746. doi: 10.1002/mds.22488 [DOI] [PubMed] [Google Scholar]
  • 48.Kasanuki K, Josephs KA, Ferman TJ, et al. Diffuse Lewy body disease manifesting as corticobasal syndrome A rare form of Lewy body disease. Neurology. 2018;91(3):E268–E279. doi: 10.1212/WNL.0000000000005828 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49.Armstrong MJ, Litvan I, Lang AE, et al. Criteria for the diagnosis of corticobasal degeneration. Neurology. 2013;80(5):496–503. doi: 10.1212/WNL.0b013e31827f0fd1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50.Ossenkoppele R, Jansen WJ, Rabinovici GD, et al. Prevalence of amyloid PET positivity in dementia syndromes: a meta-analysis. JAMA. 2015;313(19):1939–1949. doi: 10.1001/jama.2015.4669 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

eTable 1
NIHMS2152182-supplement-eTable_1.pdf (88.3KB, pdf)
eTable 2
NIHMS2152182-supplement-eTable_2.pdf (70.8KB, pdf)
eTable 3
NIHMS2152182-supplement-eTable_3.pdf (166.6KB, pdf)
eTable 4
NIHMS2152182-supplement-eTable_4.pdf (96.3KB, pdf)
eTable 5
NIHMS2152182-supplement-eTable_5.pdf (184.4KB, pdf)

Data Availability Statement

ADNI database and diagnostic information can be requested at ida.loni.usc.edu. NACC database and diagnostic details can be requested at docs.naccdata.org. PDBP database can be requested at pdbp.ninds.nih.gov.

RESOURCES