Can Deep Learning Unlock Hidden Perfusion Information from Existing Brain MRI Scans?

Validating Synthetic Data-Driven IVIM Perfusion Estimation from Sparse Multi-Shell Diffusion MRI for Alzheimer’s Disease Research

Faye McKenna, PhD

Assistant Professor, Department of Radiology

Albert Einstein College of Medicine

The Neuroimaging Research Lab

1. Overview

Scientific Rationale. Cerebrovascular dysfunction is increasingly recognized as a core driver of Alzheimer’s disease (AD) pathophysiology, not merely a comorbidity^1,2. Intravoxel incoherent motion (IVIM) diffusion MRI provides contrast-agent-free capillary microperfusion estimates, but its adoption in large-scale AD research has been blocked by the requirement for dedicated acquisition protocols with many low b-value measurements³. Thousands of participants across major neuroimaging initiatives (ADNI, UK Biobank, ABCD) have already been scanned with multi-shell diffusion protocols containing latent perfusion information that could be extracted using modern deep learning approaches⁴. This project will rigorously test whether that extraction is feasible—a no-risk, high-reward question with immediate translational impact for vascular neuroimaging research.

Funding and Facilities. This project is entirely computational and requires no new data acquisition, wet-lab resources, or participant recruitment. All work will be performed using existing high-performance computing infrastructure, which houses a dedicated GPU computing cluster. The PI has active ADNI data access through the LONI IDA platform for the subsequent clinical application phase. No IRB approval is required for this computational validation study.

Timeline. 6 months. Month 1: synthetic data generation pipeline. Months 2–3: SDD-IVIM network training and Monte Carlo benchmarking. Month 4: systematic comparison across methods and parameter regimes. Month 5: manuscript preparation and code release. Month 6: preprint, preregistration update, and submission.

Data Sharing Commitment. All analysis code, trained network weights, synthetic data generation pipelines, and Monte Carlo simulation results will be released publicly on GitHub under an open-source license. A preprint will be uploaded to ResearchHub and bioRxiv. The study protocol is preregistered on ResearchHub as part of this RFP submission.

2. Introduction

The Cerebrovascular Crisis in Alzheimer’s Disease

AD affects approximately 55 million people worldwide, with cerebrovascular dysfunction now understood to be among the earliest detectable pathophysiological events in the disease cascade—often preceding amyloid deposition by years to decades^1,5,6. Vascular contributions to cognitive impairment and dementia (VCID) encompass blood-brain barrier (BBB) breakdown, cerebral hypoperfusion, impaired glymphatic clearance, and small vessel disease, all of which accelerate and interact with canonical amyloid and tau proteinopathy^7,8. The COVID-19 pandemic has further heightened urgency around cerebrovascular neurodegeneration, as SARS-CoV-2 infection causes persistent endothelial dysfunction and capillary damage linked to accelerated cognitive decline^9,10. Developing sensitive, non-invasive biomarkers of cerebrovascular integrity is therefore a critical priority for both understanding AD pathophysiology and identifying patients who may benefit from emerging vascular-targeted therapies¹¹.

IVIM: A Hidden Perfusion Signal in Diffusion MRI

The IVIM model, introduced by Le Bihan et al. (1986)³, decomposes the diffusion-weighted MRI signal into tissue water diffusion (D) and pseudo-diffusion from blood flowing through randomly oriented capillaries (D*, with volume fraction f). The perfusion fraction f—the proportion of the voxel occupied by capillary blood flow—is the most clinically relevant parameter, correlating with capillary blood volume and regional microperfusion^12,13. Unlike arterial spin labeling (ASL), which measures arteriolar blood flow and requires labeling plane optimization, IVIM captures perfusion at the capillary level where amyloid-β plaques deposit and cause constriction¹⁴. Our preliminary data show that IVIM-derived perfusion fraction (PF) is significantly decreased in AD and more strongly associated with white matter hyperintensity burden than ASL-derived cerebral blood flow (CBF) (see Section 4, Pilot Data).

However, conventional IVIM requires dense low b-value sampling (8–16 b-values from 0–900 s/mm²), demanding dedicated scan protocols that are rarely included in large-scale studies. This has created a paradox: the datasets most valuable for studying cerebrovascular dysfunction in AD—ADNI (N > 1,000), UK Biobank (N > 40,000), ABCD (N > 10,000)—all acquired multi-shell diffusion MRI but none included IVIM-optimized b-value sampling.

The Key Innovation: Extracting IVIM from Sparse Multi-Shell Data

Recent methodological advances suggest this barrier can be overcome. Three lines of evidence indicate that meaningful perfusion fraction can be estimated from as few as 3–4 b-values:

1. Asymptotic fitting: Conklin and Federau (2016) showed that fitting a monoexponential to high b-values (≥500 s/mm²), extrapolating to b = 0, and comparing with the measured b = 0 signal yields perfusion maps comparable to full biexponential fitting¹⁵. Meeus et al. (2018) validated this specifically with b = 500 as the critical inflection point¹⁶.

2. Bayesian shrinkage priors: Bayesian approaches with data-driven priors require fewer b-values and produce more robust parameter maps with fewer outliers^17,18.

3. Synthetic data-driven deep learning (SDD-IVIM): Wang et al. (2025) trained a U-Net on synthetic IVIM parametric maps modulated with brain texture, outperforming five state-of-the-art methods with superior noise resistance⁴. Kaandorp et al. (2024) further showed that incorporating spatial information via attention models improves IVIM estimation from limited data¹⁹.

The ADNI-3/4 Advanced diffusion protocol (b = 0, 500, 1000, 2000 s/mm²)—already acquired in hundreds of well-characterized AD participants with amyloid PET, tau PET, and cognitive data—is an ideal test case. The b = 500 shell retains residual pseudo-diffusion signal, and the b = 2000 shell enables kurtosis-corrected tissue diffusion estimation for improved back-extrapolation. But no one has systematically tested whether these methods actually work at the specific b-value configuration, SNR levels, and perfusion fraction ranges relevant to the aging brain. That is the question this project will answer.

Research Questions and Hypotheses

This 6-month project addresses a single focused question: can clinically meaningful IVIM perfusion fraction be accurately estimated from the sparse b-value scheme used in ADNI-3/4 and other major multi-shell neuroimaging datasets?

H1: SDD-IVIM will estimate perfusion fraction with acceptable accuracy (bias < 15% relative to ground truth) and precision (coefficient of variation [CoV] < 25%) at brain-realistic SNR levels (≥40) for gray matter perfusion fraction values (f = 0.05–0.15).

H2: SDD-IVIM will outperform asymptotic fitting and Bayesian estimation in terms of mean absolute error and structural similarity at low-to-moderate SNR (20–40), where noise-driven bias is greatest.

H3: All methods will show degraded accuracy for f < 0.05 (white matter range) and at SNR < 30, establishing quantitative lower bounds for clinical applicability.

H4: The b = 2000 shell will significantly improve perfusion fraction estimation compared to a 3-b-value scheme (b = 0, 500, 1000) alone, by enabling kurtosis-corrected tissue diffusion estimation.

Impact Statement

If successful, this project will establish a new paradigm for cerebrovascular neuroimaging: retrospective extraction of capillary perfusion information from existing datasets that were never designed for IVIM analysis. This would immediately unlock perfusion fraction estimation for thousands of participants across ADNI, UK Biobank, ABCD, and other multi-shell datasets—without a single additional minute of scan time. Even if the method fails at certain parameter ranges, rigorously characterizing those limits is itself a valuable contribution that will guide future acquisition protocol design. This project constitutes the methodological validation phase (Aim 1) of a larger registered report on cerebrovascular dysfunction in AD (preregistered, Stage 1), with Aims 2–3 applying the validated pipeline to clinical ADNI data pending the outcome of this validation.

Figure 1. Study concept. Left: conventional IVIM requires dedicated dense low b-value acquisitions rarely included in large-scale studies. Right: the proposed approach extracts perfusion fraction from sparse multi-shell data (e.g., ADNI-3/4 Advanced) using deep learning, asymptotic fitting, and Bayesian methods, validated by Monte Carlo simulation (this 6-month project). If successful, perfusion estimation is immediately unlocked for thousands of existing participants across major neuroimaging datasets.

Relationship to Ongoing Work

This 6-month project is the first phase of a larger preregistered study on cerebrovascular dysfunction across the AD continuum. The parent project (Stage 1 Registered Report, preregistered: 10.55277/researchhub.tzuuym9h) includes three aims: Aim 1 (this proposal) validates IVIM estimation methods computationally; Aim 2 applies the validated pipeline to characterize IVIM perfusion fraction, free water, and ASL-CBF across cognitively normal, MCI, and AD groups in ADNI-3/4 (N ≈ 500–800); Aim 3 tests whether IVIM-f provides unique predictive value for AD biomarker burden beyond ASL-CBF via hierarchical regression. Aims 2–3 are contingent on the outcome of this validation and will be pursued through NIH funding mechanisms (K Award, R03). Separately, the PI is conducting IVIM-FWI analyses in two local cohorts—the Einstein Aging Study (EAS; N = 240) and a U19 Drug Discovery for Aging study (N = 200)—which use different b-value schemes and will provide independent clinical validation (10.55277/researchhub.85apbvly).

3. Methods

Participants/Sample

This is a purely computational study. No human participants will be recruited or scanned. The “sample” consists of 40,000+ synthetic diffusion-weighted MRI voxels generated by Monte Carlo simulation, plus brain-texture-modulated synthetic parametric maps for deep learning training.

Simulation Parameter Space. Synthetic data will span the physiologically plausible range of IVIM parameters in aging and AD brain tissue, informed by published values and our preliminary EAS data:

Parameter	Range	Distribution	Source
f (perfusion fraction)	0.02 – 0.20	Uniform	Le Bihan 2019; EAS pilot
D (tissue diffusion)	0.5 – 1.5 × 10⁻³ mm²/s	Gaussian (μ=0.8, σ=0.2)	Literature consensus
D* (pseudo-diffusion)	5 – 25 × 10⁻³ mm²/s	Log-normal	Federau 2017
SNR (at b=0)	20, 30, 40, 60, 80	Fixed levels	ADNI QC reports
Noise model	Rician	—	Standard MRI physics

Sample Size Justification. For each of 5 SNR levels, 10,000 voxels will be simulated across the parameter space (50,000 total), providing >99% power to detect a 5% difference in mean absolute error between methods via paired t-test (α = 0.01). For the deep learning component, 50,000 synthetic brain-texture-modulated parametric map sets will be generated for training, following Wang et al. (2025). This sample size was chosen to match the published SDD-IVIM validation protocol while exceeding it in the range of noise conditions tested.

Materials and Procedures

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 (circles/squares). The shaded region between solid (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 (ΔS ≈ 0.0002), but the asymptotic/deep learning approach works by fitting tissue diffusion (D) from high b-values and back-extrapolating to b = 0, where the perfusion offset (ΔS = f = 0.10) is fully measurable. Blue ticks show conventional IVIM b-values that ADNI lacks.

B-Value Configurations Under Test

The primary b-value scheme is the ADNI-3/4 Advanced protocol: b = 0, 500, 1000, 2000 s/mm². To test H4 (added value of b = 2000), we will also evaluate a reduced 3-b-value scheme (b = 0, 500, 1000) and a comparison “ideal” scheme with dense low b-values (b = 0, 50, 100, 200, 500, 1000, 2000) as an upper-bound reference.

Synthetic Data Generation

Step 1: Voxelwise Monte Carlo simulations. For each parameter combination and SNR level, synthesize diffusion-weighted signal using the IVIM bi-exponential forward model: S(b) = S₀ [(1–f) exp(–bD) + f exp(–bD*)]. Add Rician noise at specified SNR. Repeat 100 times per condition to estimate bias and variance.

Step 2: Brain-texture-modulated parametric maps. Following Wang et al. (2025), generate synthetic f, D, and D* parametric maps by sampling from physiological distributions and modulating with spatial texture extracted from real brain MRI (publicly available MNI templates). Simulate multi-b-value volumes from these maps with spatially varying Rician noise. These serve as training data for the SDD-IVIM network.

IVIM Estimation Methods

Three methods will be implemented, benchmarked, and compared:

1. Asymptotic/segmented fitting (reference method): Fit monoexponential to b = [500, 1000, 2000] to estimate D; extrapolate to b = 0; compute f = 1 – S₀_extrapolated / S(b=0)_measured. For the kurtosis-corrected variant, fit S(b) = S₀ exp(–bD + b²D²K/6) to high b-values before extrapolation. Implementation: scipy.optimize.curve_fit with bounded parameters.

2. Bayesian fitting with shrinkage prior: Full bi-exponential model with log-normal shrinkage prior for D* and beta prior for f, following While (2017). Implementation: custom PyMC3/NumPyro sampler. 5,000 MCMC iterations per voxel with 1,000 burn-in.

3. SDD-IVIM (primary method): 2D U-Net trained on brain-texture-modulated synthetic parametric maps, following Wang et al. (2025) adapted for the ADNI 4-b-value scheme. Architecture: encoder-decoder with skip connections, 4 encoding/decoding levels, batch normalization, ReLU activation. Loss: combined L1 + SSIM loss. Training: Adam optimizer, learning rate 1×10⁻⁴, 200 epochs, batch size 16. 5-fold cross-validation on held-out synthetic data. Implementation: PyTorch.

Equipment and Computing

All computation will be performed on the MRRC GPU cluster (NVIDIA A100 GPUs) for deep learning training and a shared HPC cluster for Monte Carlo simulations. Software environment: Python 3.10+, PyTorch 2.0+, DIPY, scipy, NumPy, nibabel. Total estimated compute: ~200 GPU-hours for SDD-IVIM training, ~50 CPU-hours for Monte Carlo simulations.

Planned Analyses

Primary Performance Metrics

For each estimation method, b-value scheme, and SNR level, we will compute:

• Mean absolute error (MAE) of estimated f relative to ground truth

• Relative bias: (f_estimated – f_true) / f_true, averaged across parameter space

• Coefficient of variation (CoV): standard deviation / mean of estimates across 100 noise realizations per condition

• Structural similarity index (SSIM) between estimated and ground truth parametric maps (for SDD-IVIM)

Hypothesis-Specific Analyses

H1 test: For SDD-IVIM at the ADNI b-value scheme, tabulate bias and CoV across the full f × SNR grid. Declare acceptable IF bias < 15% AND CoV < 25% at SNR ≥ 40 for f ≥ 0.05. Report exact bias and CoV values for all conditions.

H2 test: Paired t-tests comparing MAE of SDD-IVIM vs. asymptotic fitting and SDD-IVIM vs. Bayesian fitting at each SNR level (20, 30, 40, 60, 80), with Bonferroni correction for 10 comparisons. Report Cohen’s d effect sizes. IF SDD-IVIM MAE is significantly lower at SNR ≤ 40, THEN H2 is supported.

H3 test: Plot bias and CoV as a function of true f for each SNR level. Identify the f threshold below which CoV exceeds 25% for each method. IF all methods show CoV > 25% for f < 0.05 at SNR < 30, THEN H3 is supported. Report exact thresholds.

H4 test: Compare MAE between the 4-b-value ADNI scheme (b = 0, 500, 1000, 2000) and the 3-b-value reduced scheme (b = 0, 500, 1000) via paired t-test at each SNR level. IF the 4-b-value scheme shows significantly lower MAE (p < 0.05 after Bonferroni correction), THEN H4 is supported. Report the percentage improvement.

Contingent Analyses (IF-THEN)

IF SDD-IVIM fails H1 criteria at all SNR levels: investigate whether a transfer learning approach (pre-training on dense b-value synthetic data, fine-tuning on sparse) improves performance.

IF asymptotic fitting unexpectedly matches or outperforms SDD-IVIM: report this as a positive finding favoring simplicity, and recommend the asymptotic method for clinical adoption given its lower computational requirements.

IF the kurtosis-corrected asymptotic variant substantially outperforms the standard version: recommend that future sparse-IVIM studies include at least one high b-value (≥2000) for kurtosis correction.

Ethics and Data Management

Ethical Approval: Not required. This study is entirely computational, using only synthetic data and publicly available brain templates (MNI152). No human data are processed in this phase.

Data Sharing and Archiving: All code, trained model weights, synthetic data generation scripts, and full Monte Carlo result tables will be deposited on GitHub (public repository) and archived on Zenodo with a DOI upon manuscript submission. The preprint will be posted on ResearchHub and bioRxiv. Raw simulation outputs (estimated ~5 GB) will be archived on Zenodo.

4. Pilot Data (Optional)

Although this project is a computational validation study, its clinical motivation is grounded in preliminary IVIM and ASL perfusion data from two independent aging cohorts analyzed by the PI.

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 CBF were computed using the Rydhog et al. (2017) multi-shell estimation technique²⁰ implemented in DIPY, with FreeSurfer Desikan-Killiany atlas parcellation. Three findings directly inform this proposal:

1. IVIM-PF was significantly decreased in AD participants relative to healthy controls, with additional reductions in AD participants with a history of COVID-19.

2. IVIM-PF showed stronger regional associations with white matter hyperintensity burden compared to ASL-CBF, consistent with IVIM’s sensitivity to capillary-level pathology where amyloid-β causes constriction.

3. IVIM-PF demonstrated a coefficient of variation < 10% within each diagnostic group for > 50% of 54 brain regions examined, indicating adequate measurement reliability.

U19 Drug Discovery for Aging Study

Ongoing data collection (target N = 200) using b = 0, 500, 1000, 2000 s/mm²—identical to the ADNI-3/4 Advanced protocol targeted in this proposal. This dataset will serve as a direct bridge for validating sparse-IVIM estimates against full biexponential IVIM once additional low b-value acquisitions (b = 50, 100, 250 s/mm², funded by a Research Hub and Radiology seed grant) are completed.

Relevance to Current Proposal

The EAS pilot data establish that IVIM perfusion fraction detects meaningful cerebrovascular dysfunction in AD. The open question—addressed by this proposal—is whether comparable measurements can be obtained from the sparser b-value distributions used in ADNI and other large-scale studies. The EAS b = 300 shell is within the conventional IVIM-sensitive range; the ADNI b = 500 shell retains only residual pseudo-diffusion signal, making estimation substantially more challenging and requiring the computational methods validated here.

5. Budget

Item	Cost (USD/RSC)	Justification
GPU compute (cloud burst)	$2,500	~200 GPU-hrs A100 for SDD-IVIM training + hyperparameter sweeps; supplements on-site cluster for peak demand
Personnel: graduate student (10% effort, 3 months)	$3,500	Assist with simulation pipeline coding, data generation, and quality control
Software and tools	$500	Cloud storage for simulation outputs (~5 GB Zenodo archival), code hosting
Open-access publication fees	$1,500	Preprint hosting, potential APC for open-access journal submission
Conference presentation	$2,000	Travel to present results at ISMRM or OHBM (registration + travel)
TOTAL	$10,000

Note: Budget items are estimates and may be adjusted based on results. The PI salary is supported by institutional funds and is not charged to this project. On-site MRRC GPU cluster time is available at no cost for baseline computation; cloud burst funds cover peak-demand overflow during intensive training runs.

6. References

1. 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).

2. 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).

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

4. 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).

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

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

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

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

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

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

11. Le Bihan, D. What can we see with IVIM MRI? NeuroImage 187, 56–67 (2019).

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

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

14. Madsen, L. S. et al. Capillary dysfunction correlates with cortical amyloid load in early Alzheimer’s disease. Neurobiol. Aging 123, 1–10 (2023).

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

16. Meeus, E. M. et al. Rapid measurement of intravoxel incoherent motion (IVIM) derived perfusion fraction for clinical MRI. Magn. Reson. Mater. Phys. 31, 269–283 (2018).

17. 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).

18. Gustafsson, O. et al. Impact of prior distributions and central tendency measures on Bayesian intravoxel incoherent motion model fitting. Magn. Reson. Med. 79, 1485–1491 (2018).

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

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. 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).

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

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

Can Deep Learning Unlock Hidden Perfusion Information from Existing Brain MRI Scans?

Can Deep Learning Unlock Hidden Perfusion Information from Existing Brain MRI Scans?

Funding Progress

Funders

Related Funding

Request for Proposals - Vascular Biology

Nonprofit Recipient

Topics

DOI