Capillary Microperfusion Across the Alzheimer’s Continuum: Extracting IVIM Perfusion from Multi-Shell Diffusion MRI in ADNI

Faye McKenna, PhD

Department of Radiology, Albert Einstein College of Medicine

The Neuroimaging Research Lab

NY, USA

1. Overview

Background: Cerebrovascular dysfunction is increasingly recognized as a core pathophysiological mechanism in Alzheimer’s disease (AD), contributing to blood-brain barrier breakdown, impaired brain waste clearance, and white matter degeneration. Intravoxel incoherent motion (IVIM) diffusion MRI provides a contrast-agent-free measure of tissue microperfusion through estimation of the perfusion fraction (f), but clinical adoption has been limited by the requirement for dedicated acquisition protocols with dense low b-value sampling. Multi-shell diffusion protocols with sparse b-value distributions (e.g., b = 0, 500, 1000, 2000 s/mm²), now standard in large-scale aging and AD research studies, were not designed for IVIM analysis yet contain latent perfusion information that can be extracted using modern computational approaches.

Objectives: This registered report proposes a multi-modal neuroimaging investigation of cerebrovascular dysfunction across the AD continuum using the ADNI-3/4 Advanced diffusion dataset. The primary methodological aim is to validate synthetic data-driven deep learning (SDD-IVIM) and asymptotic fitting methods for estimating IVIM perfusion fraction from sparse multi-shell diffusion acquisitions. The primary scientific aim is to characterize the relationships among IVIM-derived perfusion fraction, free water fraction, ASL-derived cerebral blood flow, and AD biomarkers (amyloid PET, tau PET, cognition) across cognitively normal, mild cognitive impairment, and AD diagnostic groups.

Methods: In Aim 1, we will develop and validate SDD-IVIM parameter estimation from the ADNI-3/4 Advanced b-value scheme (b = 0, 500, 1000, 2000 s/mm²) using Monte Carlo simulations with brain-realistic ground truth parameters, benchmarking against asymptotic/segmented fitting and Bayesian approaches. In Aim 2, we will apply the validated pipeline to ADNI-3/4 Advanced participants (estimated N ≈ 500–800 with complete multi-modal data) to quantify IVIM perfusion fraction, free water fraction, and ASL-CBF across the AD spectrum. In Aim 3, we will test whether IVIM-derived perfusion fraction provides complementary information to ASL-CBF in predicting AD biomarker burden, cognitive decline, and diagnostic group membership, with particular attention to regional patterns in AD-vulnerable territories.

Hypotheses: We hypothesize that: (1) SDD-IVIM will yield perfusion fraction estimates with acceptable bias (<15%) and precision (coefficient of variation <25%) at brain-realistic SNR levels; (2) IVIM perfusion fraction will be significantly reduced in AD and MCI relative to cognitively normal controls, with the largest effects in temporoparietal and posterior cingulate regions; (3) reduced perfusion fraction will correlate with increased amyloid and tau burden, reduced cognitive performance, and increased free water fraction; and (4) IVIM perfusion fraction will account for unique variance in AD biomarker burden beyond ASL-CBF alone, supporting its role as a complementary cerebrovascular biomarker.

Significance: This study will establish whether clinically meaningful IVIM perfusion information can be retrospectively extracted from existing large-scale multi-shell diffusion datasets that were never designed for IVIM analysis. A positive finding would immediately unlock perfusion fraction estimation for thousands of participants across ADNI, UK Biobank, ABCD, and other major neuroimaging initiatives without additional scan time, representing a substantial translational advance for cerebrovascular neuroimaging research in AD.

2. Introduction

Research Question: Background, Importance, and Relevance

Figure 1. Study overview. The three-aim investigation flows from ADNI-3/4 Advanced multi-modal data through: Aim 1, methodological validation of IVIM estimation methods via Monte Carlo simulation (months 1–6); Aim 2, clinical characterization of IVIM-f, free water, and ASL-CBF across the AD spectrum (months 5–12); and Aim 3, multivariate integration testing whether IVIM-f provides unique predictive value for AD biomarkers beyond ASL-CBF and free water (months 10–16). Primary hypotheses (H1–H4) are mapped to their respective aims. Bottom bar shows the translational impact if validated.

Alzheimer’s disease (AD) has traditionally been conceptualized as a disorder of aberrant protein aggregation, centered on amyloid-β plaques and neurofibrillary tau tangles ⁹. However, accumulating evidence reveals that cerebrovascular dysfunction is not merely a comorbidity but a core pathophysiological contributor to AD progression, often preceding detectable amyloid deposition by years to decades ^8,21. Vascular contributions to cognitive impairment and dementia (VCID) encompass blood-brain barrier (BBB) breakdown, cerebral hypoperfusion, impaired glymphatic clearance, small vessel disease, and white matter degeneration—all of which accelerate and interact with canonical AD proteinopathy ^17,26.

The COVID-19 pandemic has heightened urgency around cerebrovascular contributions to neurodegeneration. SARS-CoV-2 infection is associated with endothelial dysfunction, microvascular thrombosis, and BBB disruption ¹⁶, and post-acute COVID-19 syndrome (“long COVID”) frequently involves persistent cognitive impairment that shares features with early AD ⁴. Understanding how cerebrovascular dysfunction contributes to neurodegeneration—and developing sensitive biomarkers to detect it early—is thus a clinical priority extending well beyond the traditional AD population.

Neuroimaging offers powerful tools for quantifying cerebrovascular integrity in vivo. Arterial spin labeling (ASL) MRI provides non-invasive measurement of cerebral blood flow (CBF), and regional hypoperfusion patterns on ASL are well-characterized in AD ¹. Free water imaging (FWI), derived from multi-shell diffusion MRI, quantifies the extracellular free water fraction as a putative marker of neuroinflammation, edema, and BBB breakdown ¹⁹. However, a third source of cerebrovascular information—intravoxel incoherent motion (IVIM) perfusion—has been largely underutilized in AD research, despite its unique capacity to provide contrast-agent-free microperfusion estimates at the capillary level.

The IVIM Model and Its Clinical Potential

The IVIM model, introduced by Le Bihan et al.¹², decomposes the diffusion-weighted MRI signal into two compartments: true tissue water diffusion (characterized by the diffusion coefficient D) and pseudo-diffusion from incoherent blood flow in randomly oriented capillaries (characterized by the pseudo-diffusion coefficient D* and the perfusion fraction f). The perfusion fraction f represents the volume fraction of the voxel occupied by flowing blood in the capillary network and is the most clinically relevant IVIM parameter, correlating with capillary blood volume and regional perfusion ⁵.

IVIM perfusion has been validated against dynamic susceptibility contrast (DSC) MRI in brain tumors and stroke ^5,6 and offers several advantages over ASL-CBF: it does not require labeling plane optimization, is less susceptible to transit time artifacts, measures microperfusion at the capillary rather than arteriolar level, and can be derived from diffusion-weighted data that are acquired for other purposes (e.g., DTI, NODDI, tractography). However, conventional IVIM analysis requires dense low b-value sampling (typically 8–16 b-values ranging from 0 to 800–900 s/mm²), which has limited its adoption in large-scale neuroimaging studies where scan time is at a premium.

Extracting IVIM from Sparse Multi-Shell Acquisitions: A Methodological Innovation

The ADNI-3/4 Advanced diffusion protocol acquires data at b = 0, 500, 1000, and 2000 s/mm²—a multi-shell scheme designed for DTI, DKI, and NODDI analysis, not for IVIM. Crucially, however, recent methodological advances demonstrate that clinically meaningful IVIM perfusion fraction can be estimated from such sparse b-value distributions. These approaches fall into three categories:

1. Asymptotic/segmented fitting: Conklin and Federau³ demonstrated that a simplified linear fitting approach—fitting a monoexponential to high b-values (b ≥ 200–500 s/mm²), extrapolating to b = 0, and comparing with the measured b = 0 signal to estimate f—produces perfusion maps qualitatively similar to full biexponential fitting. Meeus et al.¹⁵ specifically validated that a 3-b-value protocol including b = 500 s/mm² accurately captures perfusion fraction in brain tissue, identifying b = 500 as the critical inflection point that retains residual pseudo-diffusion signal while providing adequate tissue diffusion weighting.

2. Bayesian fitting with shrinkage priors: Bayesian approaches using data-driven or shrinkage priors require fewer low b-values than conventional least-squares estimation, yielding more robust parameter maps with fewer outliers and lower relative error ^7,24. These methods are particularly advantageous for brain IVIM where the perfusion fraction is small (~5%) and D* estimation is inherently noisy.

3. Synthetic data-driven deep learning (SDD-IVIM): Wang et al.²³ introduced a method that generates synthetic IVIM parametric maps sampled from physiologically plausible distributions, modulates them with human brain texture, and uses these to train a U-Net for supervised IVIM parameter mapping. This approach outperformed five state-of-the-art methods in phantom and clinical data, with superior noise resistance—exactly the properties needed for sparse b-value brain data. Additional methods incorporating spatial information through attention models ¹⁰ and physics-guided self-supervised networks ¹⁸ further demonstrate the feasibility of extracting IVIM parameters from limited data.

The availability of b = 500 s/mm² in the ADNI-3/4 Advanced protocol is critical: this b-value captures residual pseudo-diffusion effects while beginning tissue diffusion weighting, enabling the back-extrapolation to b = 0 that is the basis for perfusion fraction estimation. The additional b = 2000 shell provides kurtosis information that improves the accuracy of the tissue diffusion estimate, enhancing f estimation via better back-extrapolation.

Figure 2. Simulated IVIM signal decay for gray matter (f = 0.10, teal) and white matter (f = 0.04, orange). (A) Full b-value range with ADNI-3/4 sample points marked. The shaded region between solid (full IVIM) and dashed (tissue-only) curves represents the perfusion contribution. (B) Zoomed low b-value region. The red line marks b = 500 s/mm², the lowest non-zero ADNI shell. By b = 500, direct pseudo-diffusion signal is nearly attenuated, but the asymptotic approach fits tissue diffusion (D) from high b-values and back-extrapolates to b = 0, where the perfusion offset (ΔS = f) is fully measurable. Blue ticks show conventional IVIM b-values absent from the ADNI protocol.

Present Study: Aims and Innovation

This registered report proposes a three-aim investigation combining methodological validation with clinical application:

• Aim 1 (Methodological Validation): Develop and validate SDD-IVIM and asymptotic fitting pipelines for the ADNI-3/4 Advanced b-value scheme using Monte Carlo simulations and comparison with ASL-CBF as an independent perfusion reference.

• Aim 2 (Cross-Sectional Characterization): Apply the validated pipeline to characterize IVIM perfusion fraction, free water fraction, and ASL-CBF across the AD continuum in ADNI-3/4 participants, integrated with amyloid PET, tau PET, and cognitive data.

• Aim 3 (Multivariate Integration): Determine whether IVIM perfusion fraction contributes unique variance in predicting AD biomarker burden and cognitive decline beyond ASL-CBF and free water fraction, supporting its role as a complementary cerebrovascular biomarker.

The primary innovation is demonstrating that IVIM-derived perfusion information can be retrospectively extracted from multi-shell diffusion datasets acquired for other purposes. A positive finding would immediately unlock perfusion fraction estimation for the entire ADNI-3/4 Advanced cohort (~1,000+ participants with full AD biomarker characterization), with direct applicability to other large-scale multi-shell datasets including UK Biobank, ABCD, and numerous disease-focused Connectome projects.

Hypotheses

Primary Hypotheses:

1. H1 (Methodological): SDD-IVIM perfusion fraction estimates derived from the ADNI-3/4 b-value scheme (b = 0, 500, 1000, 2000 s/mm²) will demonstrate acceptable accuracy (bias < 15% relative to ground truth) and precision (coefficient of variation < 25%) at brain-realistic SNR (40–80) for tissue-level perfusion fraction values (f = 0.03–0.15).

2. H2 (Clinical): IVIM perfusion fraction will be significantly reduced in MCI and AD participants relative to cognitively normal (CN) controls, with the largest group differences in temporoparietal cortex, posterior cingulate/precuneus, and medial temporal regions (effect size Cohen’s d ≥ 0.40 for AD vs. CN).

3. H3 (Biomarker Integration): Reduced IVIM perfusion fraction will correlate significantly with increased amyloid PET burden (SUVR), increased tau PET burden (in Braak-stage ROIs), and reduced cognitive performance (composite memory and executive function scores), after adjusting for age, sex, and APOE ε4 status.

4. H4 (Complementary Value): IVIM perfusion fraction will account for significant unique variance (ΔR² ≥ 0.02) in predicting AD biomarker burden and cognitive scores beyond what is explained by ASL-CBF and free water fraction alone, as assessed by hierarchical regression and model comparison.

Secondary Hypotheses:

1. H5: IVIM perfusion fraction and ASL-CBF will show moderate spatial correlation (r = 0.3–0.6) at the regional level, with divergence reflecting the distinct vascular compartments measured (capillary microperfusion vs. arteriolar blood flow).

2. H6: Increased free water fraction will be associated with decreased IVIM perfusion fraction, consistent with a model in which BBB breakdown (elevated FW) co-occurs with capillary hypoperfusion (reduced f).

3. H7: APOE ε4 carriers will show greater IVIM perfusion deficits relative to non-carriers, independent of amyloid status, reflecting the known vascular effects of APOE ε4 on cerebrovascular integrity.

Exploratory Objectives:

• Characterize regional patterns of IVIM perfusion loss across Braak tau staging regions and compare with amyloid-dependent vs. amyloid-independent hypoperfusion patterns.

• Assess whether combined IVIM-f + FW + ASL-CBF improves classification accuracy for diagnostic group membership (CN vs. MCI vs. AD) relative to any single modality.

• Explore longitudinal associations between baseline IVIM perfusion fraction and subsequent cognitive decline in participants with available follow-up data.

Timeline

The study is planned for completion within 18 months. Months 1–4: ADNI data access application, data download, and preprocessing pipeline development. Months 2–6: Aim 1 (Monte Carlo validation study, SDD-IVIM network training). Months 5–12: Aim 2 (apply validated pipeline to full ADNI-3/4 Advanced cohort). Months 10–16: Aim 3 (multivariate integration analyses). Months 14–18: Manuscript preparation and Stage 2 submission.

3. Methods

Figure 3. Conceptual model of three complementary cerebrovascular biomarkers. ASL-CBF measures arteriolar-level blood flow via labeled water transit. IVIM perfusion fraction (f) measures capillary-level microperfusion via pseudo-diffusion of blood in randomly oriented capillaries—the vascular compartment where amyloid-β deposits cause constriction. Free water fraction measures extracellular water reflecting BBB breakdown and neuroinflammation. Together, these three modalities capture distinct vascular compartments (H4), and IVIM-f and FW can be co-estimated from the same multi-shell diffusion data.

Experimental Design

Dataset: ADNI-3/4 Advanced Diffusion Protocol

Data will be obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), accessed through the Laboratory of Neuro Imaging Image and Data Archive (LONI IDA; ida.loni.usc.edu). ADNI-3 and ADNI-4 include two diffusion MRI tiers: a Basic protocol (single-shell, b = 0 and 1000 s/mm²) and an Advanced protocol (multi-shell, b = 0, 500, 1000, 2000 s/mm² with 2.0 mm isotropic resolution and multi-band acceleration). Only participants scanned with the Advanced protocol will be included, as the Basic protocol lacks the b = 500 shell required for IVIM analysis.

The Advanced diffusion acquisition provides 3 non-zero b-value shells with multiple gradient directions per shell, yielding sufficient data for DTI, DKI, NODDI, and—as we propose to demonstrate—IVIM perfusion fraction estimation. The protocol also includes 3D pseudo-continuous ASL (pCASL) with multiple post-labeling delays, enabling quantitative CBF mapping for cross-validation.

Available Imaging and Biomarker Modalities:

Modality	Acquisition / Measure	Derived Outcome
Advanced Diffusion MRI	b = 0, 500, 1000, 2000; 2mm iso	IVIM-f, FW, DTI, DKI, NODDI
3D pCASL	Multi-delay pCASL	Quantitative CBF
T1 MP-RAGE	1mm isotropic	FreeSurfer parcellation, volumes
3D FLAIR	1mm isotropic	WMH segmentation
Amyloid PET	Florbetapir / Florbetaben	SUVR (centiloid)
Tau PET	Flortaucipir (subset)	Braak-stage ROI SUVR
Cognition	ADNI neuropsych battery	ADNI-Mem, ADNI-EF composites
Genetics	APOE genotyping, GWAS	ε4 carrier status
CSF (subset)	Lumipulse	Aβ42/40, p-tau181, NfL

Aim 1: Methodological Validation of Sparse-IVIM Estimation

1a. Monte Carlo Simulation Study. We will generate synthetic diffusion-weighted datasets that replicate the ADNI-3/4 Advanced acquisition (b = 0, 500, 1000, 2000 s/mm²) with known ground-truth IVIM parameters. Ground truth parameter ranges will span physiologically plausible brain values: f = 0.03–0.15 (gray and white matter), D = 0.5–1.2 × 10⁻³ mm²/s, D* = 5–20 × 10⁻³ mm²/s. Rician noise will be added at SNR levels of 20, 40, 60, and 80 (relative to b = 0 signal). For each SNR level, 10,000 voxels will be simulated across the parameter space. Three estimation methods will be compared:

• Asymptotic/segmented fitting: Monoexponential fit to b = [500, 1000, 2000] to estimate D; back-extrapolation to b = 0 to estimate f ³.

• Bayesian fitting with shrinkage prior: Implementation following While²⁴ with data-driven prior estimation from the simulated ensemble.

• SDD-IVIM (synthetic data-driven U-Net): Supervised U-Net trained on synthetic IVIM parametric maps modulated with brain texture, following Wang et al.²³. Training data will be generated using the IVIM bi-exponential forward model at the ADNI b-value scheme, with parameter distributions spanning physiological ranges and Rician noise injection.

Performance metrics: mean absolute error (MAE), relative bias, coefficient of variation (CoV), and structural similarity index (SSIM) for parametric maps.

1b. In Vivo Validation Against ASL-CBF. As no ground-truth IVIM maps exist for in vivo data, we will validate the estimated IVIM perfusion fraction against ASL-derived CBF as an independent perfusion reference. The correlation between IVIM-f and ASL-CBF will be assessed at the regional level (FreeSurfer cortical parcellation) and at the voxel level within gray matter. We do not expect perfect agreement (they measure different vascular compartments) but do expect moderate positive correlation (r = 0.3–0.6) and concordant spatial patterns of hypoperfusion.

Aim 2: Cross-Sectional Characterization of Cerebrovascular Markers in AD

The validated IVIM pipeline will be applied to all ADNI-3/4 Advanced participants meeting inclusion criteria. For each participant, we will compute: (a) voxelwise IVIM perfusion fraction (f) and tissue diffusion coefficient (D) maps; (b) free water fraction maps from multi-shell FWI ¹⁹; (c) ASL-CBF maps from the multi-delay pCASL acquisition; (d) conventional DTI metrics (FA, MD) and NODDI parameters (NDI, ODI) for reference. Regional values will be extracted using the Desikan-Killiany cortical atlas and FreeSurfer subcortical segmentation. Predefined regions of interest (ROIs) will include AD-vulnerable territories: posterior cingulate/precuneus, medial temporal (hippocampus, entorhinal cortex, parahippocampal gyrus), temporoparietal cortex (inferior parietal, supramarginal, angular gyri), lateral temporal cortex, and frontal cortex.

Aim 3: Multivariate Integration of Cerebrovascular and AD Biomarkers

Hierarchical regression models will test whether IVIM-f provides unique predictive value for AD biomarker burden beyond what is captured by ASL-CBF and free water fraction. For each primary outcome (amyloid PET SUVR, tau PET ROI SUVR, ADNI-Mem composite, ADNI-EF composite), we will build models in three steps: Step 1 includes demographics and APOE ε4 status; Step 2 adds ASL-CBF and free water fraction; Step 3 adds IVIM-f. The change in R² at Step 3 (ΔR²) is the primary test of H4. Supplementary random forest models will assess nonlinear contributions and feature importance rankings.

Inclusion and Exclusion Criteria

Inclusion: ADNI-3 or ADNI-4 participants scanned with the Advanced diffusion protocol who have (a) complete multi-shell diffusion data (b = 0, 500, 1000, 2000), (b) 3D pCASL data, (c) T1 MP-RAGE for FreeSurfer parcellation, (d) clinical diagnosis (CN, MCI, or AD), and (e) amyloid PET data. Tau PET and CSF biomarkers are available in subsets and will be analyzed as secondary outcomes.

Exclusion: Gross motion artifacts on diffusion MRI (defined as >2mm mean framewise displacement); incomplete diffusion acquisition (<80% of expected volumes); failed FreeSurfer reconstruction; clinical diagnoses other than CN, MCI, or AD (e.g., frontotemporal dementia, Lewy body dementia). Participants with vascular dementia are excluded from primary analyses but will be characterized separately as an exploratory analysis.

Sample and Statistical Power

The ADNI-3/4 database includes approximately 1,700+ participants, with an estimated 500–800 having been scanned with the Advanced diffusion protocol and meeting full multi-modal data requirements. Based on prior neuroimaging studies of AD-related perfusion deficits ^2,25, we anticipate medium-to-large effect sizes (Cohen’s d = 0.40–0.80) for group differences in perfusion measures between CN and AD groups. Power analysis (G*Power 3.1) indicates that N = 64 per group provides 80% power to detect d = 0.50 at α = 0.05 (two-tailed) for independent samples comparisons, and N = 191 total provides 80% power to detect f² = 0.04 (ΔR² = 0.02) in hierarchical regression with 8 predictors. The anticipated sample size (N ≥ 500) provides substantial power for all planned analyses, including subgroup analyses by APOE genotype and amyloid positivity status.

Covariates: Age, sex, years of education, APOE ε4 carrier status, scanner site, and total intracranial volume will be included as covariates in all group comparison and regression analyses.

Data Processing Pipeline

Diffusion MRI Preprocessing

All diffusion data will be preprocessed using a standardized pipeline: (1) denoising via MP-PCA ²²; (2) Gibbs ringing correction ¹¹; (3) susceptibility-induced distortion correction using Synb0-DISCO or reverse phase-encoding pairs where available; (4) eddy current and motion correction using FSL’s eddy with outlier replacement and slice-to-volume correction; (5) bias field correction (N4ITK). Quality control will include automated detection of signal dropout, excessive motion (mean framewise displacement > 2mm), and visual inspection of a random 10% subsample.

IVIM Perfusion Fraction Estimation

Three complementary methods will be applied, with SDD-IVIM as the primary method and asymptotic fitting as the reference method:

• Asymptotic fitting (reference): Step 1: Fit monoexponential S(b) = S₀ · exp(–b · D) to b = [500, 1000, 2000] to estimate D. Step 2: Extrapolate to b = 0 using estimated D. Step 3: f = 1 – S₀_extrapolated / S(b=0)_measured. The b = 2000 shell enables kurtosis-corrected D estimation, improving back-extrapolation accuracy.

• SDD-IVIM (primary): Train U-Net on synthetic parametric maps following Wang et al.²³ protocol, adapted for the 4-b-value ADNI scheme. Network architecture: 2D U-Net with spatial smoothness regularization. Training: 50,000 synthetic brain-texture-modulated parametric map sets with Rician noise injection at SNR 20–80. 5-fold cross-validation on held-out simulated data. Inference on real ADNI data will be slice-by-slice with subsequent 3D spatial smoothing.

• Bayesian fitting (comparison): Full biexponential model with log-normal shrinkage prior for D* and beta prior for f, following While²⁴. 5,000 MCMC iterations per voxel with 1,000 burn-in.

Free Water Imaging

Free water fraction will be estimated from the multi-shell data (b = 0, 500, 1000, 2000) using the bi-tensor model ¹⁹, separating the diffusion signal into a tissue compartment (anisotropic tensor) and an isotropic free water compartment. The free water fraction (FW) represents the volume fraction of the voxel attributed to freely diffusing extracellular water. Implementation will use DIPY’s fwDTI module.

ASL Processing

ASL data from the multi-delay pCASL acquisition will be processed using a modified version of the HCP ASL processing pipeline or Oxford ASL (FSL BASIL). Quantitative CBF maps (ml/100g/min) will be computed using the general kinetic model with multi-delay fitting to jointly estimate CBF and arterial transit time (ATT). Partial volume correction will be applied using the Luo method ¹³. CBF maps will be registered to each participant’s T1 space for ROI extraction.

Structural MRI and Biomarker Processing

T1 MP-RAGE data will be processed through FreeSurfer 7.x for cortical parcellation (Desikan-Killiany atlas), subcortical segmentation, and cortical thickness estimation. FLAIR images will be processed for WMH segmentation using the DeCarli lab’s protocol (ADNI standard). Amyloid PET SUVR will be computed using the whole cerebellum as reference region, with centiloid conversion. Tau PET (flortaucipir) SUVR will be extracted from Braak-stage ROIs (I–II: entorhinal; III–IV: hippocampus, fusiform, inferior/middle temporal; V–VI: widespread cortical).

Figure 4. Data processing and analysis pipeline. Multi-shell diffusion MRI is preprocessed (denoising, Gibbs correction, distortion correction, eddy correction, bias field correction) and processed through three parallel IVIM estimation methods (asymptotic fitting, SDD-IVIM deep learning, Bayesian shrinkage) plus free water imaging. pCASL data yield quantitative CBF maps via Oxford ASL (BASIL). T1 MP-RAGE provides FreeSurfer cortical parcellation (Desikan-Killiany atlas) for ROI extraction. Regional values from all cerebrovascular markers are integrated with amyloid PET, tau PET, and cognitive data in the statistical analysis framework (ANCOVA, partial correlations, hierarchical regression, and diagnostic classification).

Statistical Methods

Aim 1 Analyses:

• Monte Carlo: For each estimation method × SNR level, compute MAE, relative bias, CoV, and SSIM. Declare method acceptable if bias < 15% and CoV < 25% at SNR ≥ 40 for f ≥ 0.05 (gray matter range). Compare methods via paired t-tests on absolute error across parameter space.

• In vivo validation: Pearson/Spearman correlation between regional IVIM-f and ASL-CBF across Desikan-Killiany ROIs. Bland-Altman analysis after z-score normalization. Spatial concordance of hypoperfusion patterns via dice overlap of bottom-quartile regions.

Aim 2 Analyses:

• Group comparisons: ANCOVA comparing IVIM-f, FW, and ASL-CBF across diagnostic groups (CN, MCI, AD) with age, sex, site, and education as covariates. Effect sizes (Cohen’s d) and 95% CI reported. FDR correction across ROIs within each modality.

• Regional patterns: Voxelwise analyses using FSL randomise (5,000 permutations) to identify clusters of significant group differences, with threshold-free cluster enhancement (TFCE) and family-wise error correction at p < 0.05.

• Correlations with AD biomarkers: Partial correlations between each cerebrovascular marker and amyloid SUVR, tau ROI SUVR, ADNI-Mem, and ADNI-EF, controlling for age, sex, education, and APOE ε4 status.

Aim 3 Analyses:

• Hierarchical regression: For each outcome variable, compare nested models using F-test for ΔR². Report standardized betas, ΔR², and 95% CI for each predictor.

• Random forest: Variable importance rankings (mean decrease in accuracy) for IVIM-f, FW, ASL-CBF, and their interaction terms. Comparison of out-of-bag prediction accuracy across modality combinations.

• Diagnostic classification: Multinomial logistic regression and random forest classification of CN vs. MCI vs. AD using cerebrovascular features alone and combined with amyloid status. Report AUC, sensitivity, specificity with 10-fold cross-validation.

Missing Data: Participants with >20% missing ROI values (due to segmentation failures) will be excluded. For remaining missing data, multiple imputation (m = 20) will be used. Sensitivity analyses comparing complete-case and imputed results will be reported.

Multiple Comparisons: FDR correction (Benjamini-Hochberg) applied within each aim’s primary tests. Primary hypotheses (H1–H4) each correspond to a single primary test or a small family of tests corrected within aim. Exploratory analyses reported with uncorrected p-values and effect sizes.

Software: Diffusion preprocessing: FSL 6.x, MRtrix3, DIPY. IVIM estimation: Custom Python (PyTorch for SDD-IVIM, scipy.optimize for asymptotic fitting). Free water imaging: DIPY fwDTI. ASL: FSL BASIL / Oxford ASL. Statistics: R 4.3+ (lme4, randomForest, pROC) and Python (scikit-learn, statsmodels). All analysis code will be made publicly available on GitHub.

Figure 5. Hypothesized regional IVIM perfusion deficits in AD (H2). (A) Lateral view showing a priori AD-vulnerable regions: medial temporal lobe (MTL, red, predicted largest effect), temporoparietal junction (TPJ, orange), parietal cortex (gold), and frontal cortex (blue, reference/smaller effect). (B) Medial view showing Braak tau staging regions: posterior cingulate/precuneus (PCC, red, predicted largest effect), entorhinal cortex/hippocampus (EC/Hip, orange, Braak I–II), and cingulate cortex (gold, Braak V–VI). Effect sizes are hypothesized based on published ASL-CBF patterns in AD and our preliminary EAS IVIM data.

Interpreting Results

If H1 is confirmed (acceptable SDD-IVIM accuracy from sparse b-values), this will establish a new methodological paradigm: IVIM perfusion fraction can be retrospectively extracted from existing multi-shell datasets without any modification to scan protocols. This has immediate practical implications, unlocking perfusion analyses for thousands of ADNI participants and potentially for UK Biobank, ABCD, and other large-scale neuroimaging datasets. If H1 is partially confirmed (acceptable for some SNR/parameter ranges but not others), we will characterize the specific conditions under which sparse-IVIM is reliable and recommend minimum data quality criteria.

If H2–H3 are confirmed (AD-related perfusion deficits measurable by IVIM), this will demonstrate clinical relevance of sparse-IVIM in a well-characterized cohort with gold-standard AD biomarkers. If IVIM perfusion deficits parallel known ASL-CBF patterns, this provides convergent validity; if spatial patterns diverge, this may reveal complementary information about different vascular compartments affected in AD.

If H4 is confirmed (unique variance from IVIM-f beyond ASL-CBF), this would establish IVIM perfusion fraction as a distinct cerebrovascular biomarker with additive value for AD staging and prognosis. If disconfirmed, this would suggest that IVIM-f and ASL-CBF are functionally redundant in the AD context, though IVIM-f would still have value in datasets lacking ASL.

If all hypotheses are disconfirmed (SDD-IVIM fails or shows no clinical associations), this would provide rigorous negative evidence establishing the limits of sparse b-value IVIM in brain applications, which is itself a valuable contribution to the neuroimaging methods literature.

4. Pilot Data (Optional)

Although no IVIM analyses have yet been performed on ADNI data specifically, the principal investigator has conducted preliminary IVIM and ASL perfusion analyses in two independent aging and cognitive impairment cohorts that directly inform the feasibility and expected effect sizes for the proposed study.

Preliminary Dataset 1: Einstein Aging Study (EAS). Multi-shell diffusion MRI (b = 0, 300, 800, 2000 s/mm²; 1.5 mm isotropic) and single-delay pCASL data were acquired on a 3T Siemens Prisma in N = 240 participants (160 cognitively impaired [AD, Mild-AD, MCI] and 80 age-matched healthy controls; ages ≥ 70 years). IVIM-FWI perfusion fraction (PF), free water (FW), and ASL-derived cerebral blood flow (CBF) were computed using the multi-shell estimation technique ²⁰ implemented in DIPY, with FreeSurfer-based Desikan-Killiany atlas parcellation for regional analysis.

Preliminary Dataset 2: U19 Drug Discovery for Aging Study. Multi-shell diffusion MRI (b = 0, 500, 1000, 2000 s/mm²; 1.5 mm isotropic) and pCASL data are being acquired on the same 3T Prisma scanner in N = 200 participants (120 cognitively impaired individuals with a history of COVID-19 and 60 matched healthy controls; ages ≥ 60). Enrollment is ongoing with a target completion of August 2026. Seed grant funding has been secured to add low b-value IVIM acquisitions (b = 50, 100, 250 s/mm²) to the existing U19 diffusion protocol for full biexponential IVIM fitting. Notably, the U19 b-value scheme (b = 0, 500, 1000, 2000) is identical to the ADNI-3/4 Advanced protocol proposed in the present study.

Key Preliminary Findings: Analysis of the EAS cohort revealed three findings directly relevant to the present proposal: (1) IVIM perfusion fraction was significantly decreased in AD participants relative to healthy controls, with additional reductions observed in AD participants with a history of COVID-19; (2) IVIM-PF showed stronger regional associations with white matter hyperintensity burden compared to ASL-derived CBF, consistent with IVIM’s sensitivity to capillary-level microperfusion pathology where amyloid-β deposits in capillary walls cause constriction; and (3) IVIM-PF demonstrated a coefficient of variation < 10% within each diagnostic group for more than 50% of the 54 brain regions examined, indicating adequate measurement reliability for group comparison studies. These preliminary results provide direct empirical evidence that IVIM perfusion fraction detects AD-related cerebrovascular dysfunction and offers complementary information to ASL-CBF, supporting both H2 (group differences) and H4 (unique variance) of the present proposal.

Relevance to ADNI Methodology: The EAS b-value scheme (b = 0, 300, 800, 2000) includes b = 300 s/mm² in the IVIM-sensitive range, enabling conventional IVIM-FWI fitting. The ADNI-3/4 Advanced scheme (b = 0, 500, 1000, 2000) lacks dedicated low b-values below 500 s/mm², requiring the deep learning and asymptotic fitting approaches described in Aim 1. However, the b = 500 shell retains residual pseudo-diffusion signal ¹⁵, and the U19 dataset—which uses the identical b-value scheme—will serve as a direct bridge for validating sparse-IVIM estimation against the full IVIM-FWI model once low b-value acquisitions are added. Published methodological evidence further supports feasibility: Conklin and Federau³ demonstrated asymptotic fitting in brain with as few as 3–4 b-values; Wang et al.²³ showed SDD-IVIM superiority at low SNR. The PI has extensive experience with multi-shell diffusion analysis pipelines using DIPY, FSL, MRtrix3, and custom Python code, and has active ADNI data access through the LONI IDA platform.

Dissemination of Information

All analysis code (preprocessing, IVIM estimation, statistical models) will be made publicly available on GitHub under an open-source license. Trained SDD-IVIM network weights will be shared to enable direct application to other datasets with compatible b-value schemes. Derived IVIM parametric maps and regional values will be deposited in the ADNI data repository for use by the broader research community. Results will be submitted for peer-reviewed publication regardless of outcome. Preprints will be posted on bioRxiv prior to peer review.

Study Status

At the time of writing, no data analyses have been performed. The principal investigator has an approved ADNI data use agreement and access to the LONI IDA platform. Preprocessing pipeline development has not begun. The SDD-IVIM network has not been trained. All analyses described in this protocol are prospectively registered and will be conducted exactly as specified. Any deviations will be transparently reported in the Stage 2 manuscript.

5. Budget

Category	Cost (USD)	Justification
PI salary (5% effort, 18 months)	$18,000	Oversee all three aims, pipeline development, statistical analysis, manuscript preparation
Graduate student (15% effort, 12 months)	$12,000	Assist with Monte Carlo simulation pipeline, ADNI data processing, quality control
GPU cloud compute	$6,000	~300 GPU-hours A100 for SDD-IVIM training, hyperparameter optimization, and full ADNI cohort inference; supplements on-site MRRC cluster for peak demand
HPC storage and data transfer	$2,000	ADNI data download and staging (~2 TB raw diffusion), intermediate processing outputs, long-term archival
Conference travel (×2)	$5,000	Present results at ISMRM and OHBM annual meetings (registration + travel)
Open-access publication fees	$3,500	Article processing charges for registered report (Stage 2) and methods validation paper
Software and tools	$1,500	Zenodo DOI archival, cloud backup, code repository hosting
Indirect costs / contingency	$2,000	Administrative costs or reallocation to other categories as needed
TOTAL	$50,000	18-month project period

Note: Budget items are estimates and may be adjusted based on project needs. PI institutional salary and benefits are partially supported by departmental funds. On-site MRRC GPU cluster and HPC resources are available at no direct cost for baseline computation; cloud compute funds cover overflow during intensive training and full-cohort processing phases. No wet-lab supplies, participant recruitment, or MRI scan costs are required as this study uses existing ADNI data exclusively.

6. Other Required Information

Acknowledgements

Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf

Funding

Funding Source: [To be determined / Grant pending]. Principal Investigator: Faye McKenna, PhD. ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from multiple sponsors as listed at www.adni-info.org

Ethical Approval

ADNI was approved by the institutional review boards of all participating institutions. All ADNI participants provided written informed consent. The present study involves secondary analysis of existing de-identified data and does not require additional IRB approval. The principal investigator holds an approved ADNI data use agreement.

Registration

This study protocol will be registered on OSF Registries upon Stage 1 acceptance. The study does not involve a clinical trial.

Conflicts of Interest

The authors declare that they have no conflicts of interest.

4. References

1. Alsop, D. C. et al. Recommended implementation of arterial spin-labeled perfusion MRI for clinical applications. Magn. Reson. Med. 73, 102–116 (2015).

2. Binnewijzend, M. A. A. et al. Cerebral blood flow measured with 3D pseudocontinuous arterial spin-labeling MR imaging in Alzheimer disease and mild cognitive impairment. Radiology 270, 235–243 (2013).

3. Conklin, C. J. & Federau, C. Simplified model for intravoxel incoherent motion perfusion imaging of the brain. Am. J. Neuroradiol. 37, 2227–2232 (2016).

4. Douaud, G. et al. SARS-CoV-2 is associated with changes in brain structure in UK Biobank. Nature 604, 697–707 (2022).

5. Federau, C. et al. Measuring brain perfusion with intravoxel incoherent motion (IVIM): initial clinical experience. J. Magn. Reson. Imaging 39, 624–632 (2014).

6. Federau, C. Intravoxel incoherent motion MRI as a means to measure in vivo perfusion: a review of the evidence. NMR Biomed. 30, e3780 (2017).

7. Gustafsson, O., Montelius, M., Starck, G. & Ljungberg, M. Impact of prior distributions and central tendency measures on Bayesian intravoxel incoherent motion model fitting. Magn. Reson. Med. 79, 1485–1491 (2018).

8. Iturria-Medina, Y. et al. Early role of vascular dysregulation on late-onset Alzheimer’s disease based on multifactorial data-driven analysis. Nat. Commun. 7, 11934 (2016).

9. Jack, C. R. Jr et al. NIA-AA Research Framework: Toward a biological definition of Alzheimer’s disease. Alzheimers Dement. 14, 535–562 (2018).

10. Kaandorp, M. P. T. et al. Incorporating spatial information in deep learning parameter estimation with application to IVIM. Med. Image Anal. 99, 103396 (2024).

11. Kellner, E., Dhital, B., Kiselev, V. G. & Reisert, M. Gibbs-ringing artifact removal based on local subvoxel-shifts. Magn. Reson. Med. 76, 1574–1581 (2016).

12. Le Bihan, D. et al. MR imaging of intravoxel incoherent motions: application to diffusion and perfusion in neurologic disorders. Radiology 161, 401–407 (1986).

13. Luo, Y. et al. Partial volume correction for arterial spin labeling data using spatial information. Magn. Reson. Med. 72, 1800–1806 (2014).

14. McKenna, F. et al. Microstructural and microvascular alterations in psychotic spectrum disorders: a three-compartment intravoxel incoherent motion imaging and free water model. Schizophr. Bull. 49, 1542–1553 (2023).

15. Meeus, E. M., Novak, J., Dehghani, H. & Peet, A. C. Rapid measurement of intravoxel incoherent motion (IVIM) derived perfusion fraction for clinical MRI. MAGMA 31, 269–283 (2018).

16. Nalbandian, A. et al. Post-acute COVID-19 syndrome. Nat. Med. 27, 601–615 (2021).

17. Nedergaard, M. & Goldman, S. A. Glymphatic failure as a final common pathway to dementia. Science 370, 50–56 (2020).

18. Parker, C. S. et al. Self-supervised IVIM DWI parameter estimation with a physics based forward model. Magn. Reson. Imaging 83, 152–168 (2021).

19. Pasternak, O. et al. Free water elimination and mapping from diffusion MRI. Magn. Reson. Med. 62, 717–730 (2009).

20. Rydhog, A. S. et al. Separating blood and water: perfusion and free water elimination from diffusion MRI in the human brain. NeuroImage 156, 423–434 (2017).

21. Sweeney, M. D., Sagare, A. P. & Zlokovic, B. V. Blood-brain barrier breakdown in Alzheimer disease and other neurodegenerative disorders. Nat. Rev. Neurol. 14, 133–150 (2018).

22. Veraart, J. et al. Denoising of diffusion MRI using random matrix theory. NeuroImage 142, 394–406 (2016).

23. Wang, L. et al. Improved deep learning-based IVIM parameter estimation via the use of more ‘realistic’ simulated brain data. Med. Phys. 52, 651–663 (2025).

24. While, P. T. A comparative simulation study of Bayesian fitting approaches to intravoxel incoherent motion modeling in diffusion-weighted MRI. Magn. Reson. Med. 78, 2373–2387 (2017).

25. Yoshiura, T. et al. Arterial spin labelling at 3-T MR imaging for detection of individuals with Alzheimer’s disease. Eur. Radiol. 19, 2819–2825 (2009).

26. Zlokovic, B. V. Neurovascular pathways to neurodegeneration in Alzheimer’s disease and other disorders. Nat. Rev. Neurosci. 12, 723–738 (2011).

Capillary Microperfusion Across the Alzheimer’s Continuum: Extracting IVIM Perfusion from Multi-Shell Diffusion MRI in ADNI

Funding Progress

Nonprofit Recipient

Topics

DOI

Funding Progress

Nonprofit Recipient

Topics

DOI

Capillary Microperfusion Across the Alzheimer’s Continuum: Extracting IVIM Perfusion from Multi-Shell Diffusion MRI in ADNI

Funding Progress

Nonprofit Recipient

Topics

DOI