Deep-learning-enabled Brain Hemodynamic Mapping Using Resting-state fMRI

Xirui Hou, Pengfei Guo, Puyang Wang, Peiying Liu, Doris D. M. Lin, Hongli Fan, Yang Li, Zhiliang Wei, Zixuan Lin, Dengrong Jiang, Jin Jin, Catherine Kelly, Jay J. Pillai, Judy Huang, Marco C. Pinho, Binu P. Thomas, Babu G. Welch, Denise C. Park, Vishal M. Patel, Argye E. Hillis, Hanzhang Lu

Abstract

Cerebrovascular disease remains a leading cause of global mortality, making early intervention and personalized prognosis essential for effective management. While non-invasive imaging holds promise for patient stratification, current methods often lack the sensitivity required for clinical decision-making. This paper demonstrates that resting-state functional magnetic resonance imaging (rs-fMRI)—a widely available tool typically used to map neural activity—can be repurposed to map cerebral hemodynamic function. By utilizing deep learning to exploit spontaneous breathing variations as a natural "contrast media," the researchers developed a method to produce reproducible maps of cerebrovascular reactivity (CVR) and bolus arrival time (BAT). Trained against a reference standard of CO2-inhalation MRI, the deep-learning network proved effective across diverse cohorts, including healthy aging subjects and patients with Moyamoya disease, brain tumors, and stroke. This approach offers a non-invasive, highly reproducible alternative for clinical cerebrovascular imaging.

Introduction

Understanding the health of the brain's blood vessels is critical for managing conditions such as stroke, Moyamoya disease, and brain tumors. Two key parameters in this assessment are Cerebrovascular Reactivity (CVR)—the ability of blood vessels to dilate in response to stress—and Bolus Arrival Time (BAT)—the speed at which blood reaches specific brain tissues.

Historically, the "gold standard" for measuring these parameters has involved hypercapnia maneuvers, where patients must inhale CO2-enriched gas, or the administration of exogenous contrast agents. These methods are technically demanding, require specialized equipment, and are often poorly tolerated by severely ill or elderly patients. While resting-state fMRI (rs-fMRI) is non-invasive and ubiquitous in clinical settings, its use for hemodynamic mapping has been limited by a low signal-to-noise ratio (SNR). Natural breathing produces much smaller CO2 fluctuations than controlled gas inhalation, making it difficult for traditional statistical methods to extract clean, clinically useful data.

A Deep Learning Framework for Hemodynamics

To overcome the SNR limitations of rs-fMRI, the researchers developed a hierarchical deep-learning framework based on an encoder-decoder architecture (similar to a U-Net). The core innovation lies in the realization that the "noise" in rs-fMRI—specifically the fluctuations caused by natural breathing—contains latent physiological information sufficient to reconstruct high-fidelity vascular maps.

Multi-Pronged Input Strategy

Rather than feeding raw Blood-Oxygen-Level-Dependent (BOLD) time series into the network, the model utilizes a sophisticated input strategy:

  1. Primary Inputs: Parametric maps generated via traditional global-regression resting-state (GRRS) methods.
  2. Supplementary Inputs: 133 cross-correlation maps derived from the residual BOLD signal after brain parcellation.

This "supplementary encoder" is vital for accuracy in diseased brains. In pathologies like Moyamoya disease, the global BOLD signal often becomes less informative; the local residuals carry the necessary details to identify specific regional impairments.

Clinical Validation and Performance

The study utilized a diverse dataset of 232 participants from two major medical centers (UTSW and JHU), encompassing a lifespan cohort (ages 20–88) and clinical patients with various vascular and oncological conditions.

Superior Accuracy and Sensitivity

The deep-learning resting-state (DLRS) maps showed significantly higher congruency with ground-truth CO2-inhalation maps across all quantitative metrics, including Pearson correlation and Structural Similarity Index (SSIM). Beyond mere visual similarity, the DLRS maps demonstrated superior clinical sensitivity. They successfully detected physiological improvements following revascularization surgery in Moyamoya patients and showed stronger correlations with arterial stenosis grades and tumor malignancy than previous methods.

High Reproducibility

Reliability is a prerequisite for clinical adoption. In test-retest scenarios involving 67 healthy participants and 30 stroke patients, the DLRS method achieved an Intraclass Correlation Coefficient (ICC) of 0.86–0.87. This vastly outperforms traditional GRRS approaches, which typically hover between 0.38 and 0.72, suggesting that the deep-learning model is capturing stable physiological traits rather than transient noise.

Insights into Normal Aging

The model also proved sensitive to the subtle hemodynamic shifts associated with normal aging. It successfully captured the expected physiological decline in CVR and the increase in BAT (hemodynamic delay) across the lifespan cohort. This capability positions the tool not just as a diagnostic aid for acute disease, but as a potential biomarker for vascular health in the general population.

Limitations and Future Directions

While the DLRS method represents a significant leap forward, some limitations remain. Currently, the maps provide values in relative units rather than absolute physiological units (e.g., %BOLD/mmHg CO2). This makes the tool excellent for detecting regional deficits but less ideal for absolute global quantification. Additionally, while reproducibility was high in stroke patients, direct validation against CO2-inhalation "ground truth" was not possible in that cohort due to the clinical risks of the gas-challenge procedure.

Looking ahead, these hemodynamic maps could serve to "calibrate" traditional fMRI studies. By allowing researchers to separate neural activity from underlying vascular health—a long-standing challenge in neuroimaging—this deep-learning approach could refine our understanding of the functional brain in both health and disease.

Summary updated

Abstract

Cerebrovascular disease is a leading cause of death globally. Prevention and early intervention are known to be the most effective forms of its management. Non-invasive imaging methods hold great promises for early stratification, but at present lack the sensitivity for personalized prognosis. Resting-state functional magnetic resonance imaging (rs-fMRI), a powerful tool previously used for mapping neural activity, is available in most hospitals. Here we show that rs-fMRI can be used to map cerebral hemodynamic function and delineate impairment. By exploiting time variations in breathing pattern during rs-fMRI, deep learning enables reproducible mapping of cerebrovascular reactivity (CVR) and bolus arrive time (BAT) of the human brain using resting-state CO2 fluctuations as a natural 'contrast media'. The deep-learning network was trained with CVR and BAT maps obtained with a reference method of CO2-inhalation MRI, which included data from young and older healthy subjects and patients with Moyamoya disease and brain tumors. We demonstrate the performance of deep-learning cerebrovascular mapping in the detection of vascular abnormalities, evaluation of revascularization effects, and vascular alterations in normal aging. In addition, cerebrovascular maps obtained with the proposed method exhibited excellent reproducibility in both healthy volunteers and stroke patients. Deep-learning resting-state vascular imaging has the potential to become a useful tool in clinical cerebrovascular imaging.
BibTeX
@article{hou2023deep,
  author = {Xirui Hou and Pengfei Guo and Puyang Wang and Peiying Liu and Doris D. M. Lin and Hongli Fan and Yang Li and Zhiliang Wei and Zixuan Lin and Dengrong Jiang and Jin Jin and Catherine Kelly and Jay J. Pillai and Judy Huang and Marco C. Pinho and Binu P. Thomas and Babu G. Welch and Denise C. Park and Vishal M. Patel and Argye E. Hillis and Hanzhang Lu},
  title = {{Deep-learning-enabled Brain Hemodynamic Mapping Using Resting-state fMRI}},
  journal = {Npj Digit. Medicine},
  year = {2023},
  doi = {10.1038/s41746-023-00859-y},
}
Figure