Deep-learning-enabled brain hemodynamic mapping using resting-state fMRI.

Area of Science:

• Neuroimaging
• Vascular Biology
• Artificial Intelligence

Background:

• Cerebrovascular disease is a major global health concern, necessitating improved diagnostic tools.
• Current non-invasive imaging lacks sensitivity for personalized prognosis of cerebrovascular conditions.
• Resting-state functional magnetic resonance imaging (rs-fMRI) is widely available but underutilized for vascular assessment.

Purpose of the Study:

• To develop and validate a deep learning-based method for mapping cerebrovascular reactivity (CVR) and bolus arrival time (BAT) using rs-fMRI.
• To demonstrate the utility of this novel technique in detecting vascular abnormalities and assessing treatment effects.
• To establish the reproducibility of deep learning-based cerebrovascular mapping in diverse patient populations.

Main Methods:

• Utilized natural CO2 fluctuations during rs-fMRI to map cerebral hemodynamic function.
• Developed a deep learning network trained on CVR and BAT maps from a CO2-inhalation MRI reference method.
• Included data from healthy young/older adults and patients with Moyamoya disease and brain tumors.

Main Results:

• Successfully mapped CVR and BAT using rs-fMRI and deep learning, leveraging resting-state CO2 variations.
• Demonstrated the method's effectiveness in identifying vascular abnormalities, evaluating revascularization, and assessing aging-related vascular changes.
• Showcased excellent reproducibility of the deep learning-based cerebrovascular maps in healthy volunteers and stroke patients.

Conclusions:

• Deep learning analysis of rs-fMRI enables non-invasive mapping of cerebrovascular function.
• This technique offers a promising, reproducible tool for clinical cerebrovascular imaging and personalized patient management.
• Potential to enhance early diagnosis and monitoring of cerebrovascular diseases.