Reconstructing Physiological Waveforms and Mapping Cerebrovascular Reactivity Directly from fMRI Using Deep Learning

Functional magnetic resonance imaging (fMRI) is a widely used tool for mapping brain activity and connectivity, but its signals are strongly influenced by physiological fluctuations unrelated to neural activity. Low-frequency variations in respiration (RV) and cardiac rate variability (HRV) are major confounds in resting-state fMRI, yet they also provide insight into cerebrovascular dynamics. This thesis develops machine learning models to reconstruct RV and HRV waveforms directly from BOLD data, enabling their dual use for physiological confound correction and for vascular assessment. Large-scale neuroimaging datasets, the Human Connectome Project in Development (HCP-D), Young Adults (HCP-YA), and Aging (HCP-A), were used to train and validate deep learning models that estimate RV and HRV across the lifespan. These models achieved accurate waveform reconstruction without requiring external recordings. Building on this framework, reconstructed RV regressors were incorporated into a general linear model (GLM) to estimate cerebrovascular reactivity (CVR) from resting-state fMRI. For validation, a locally collected dataset acquired at Foothills Hospital, Calgary, with gas-challenge CVR measurements served as the reference standard. The results demonstrate that RV and HRV can be robustly estimated from fMRI data alone and effectively reduce low-frequency physiological confounds. Furthermore, the GLM-based application enables calibration-free CVR mapping that is consistent with conventional gas-challenge protocols. By eliminating the need for additional equipment such as respiratory belts, pulse oximeters, and specialized end-tidal monitoring systems, the proposed method reduces cost, complexity, and participant burden. Together, these findings highlight the potential of machine learning to enhance the physiological interpretability of fMRI and to provide non-invasive, scalable biomarkers of cerebrovascular health.