Abstract
Cerebrovascular reactivity (CVR) is a critical indicator of brain health, measuring the ability of cerebral blood vessels to dilate in response to vasoactive stimuli like carbon dioxide ($CO_2$). While breath-holding tasks are commonly used in fMRI to assess CVR, obtaining accurate end-tidal $CO_2$ ($P_{ET}CO_2$) measurements via nasal cannula requires high patient compliance, which is often lacking in clinical populations. This study explores a deep learning-based "rescue" pipeline that predicts $P_{ET}CO_2$ from robust respiration belt measurements (RVT) and rescales the output to standard units (%BOLD/mmHg). Our results demonstrate that this method produces CVR amplitude and delay maps with high absolute agreement to ground truth, even in cases of cerebrovascular pathology like Moyamoya disease.
Introduction
CVR mapping serves as a "stress test" for the brain's vasculature. By measuring the BOLD fMRI response to changes in arterial $CO_2$, clinicians can identify regions of impaired blood flow regulation. Standardizing these measurements into units of %BOLD/mmHg is essential for cross-subject comparisons and longitudinal tracking. However, the gold-standard measurement of $P_{ET}CO_2$ is notoriously difficult to obtain; it requires participants to breathe exclusively through their nose and perform specific exhalations. When these requirements are not met, the resulting $CO_2$ data is often too noisy for quantitative analysis.
Previous efforts have used Respiration Volume per Time (RVT) as a surrogate, but RVT lacks standard units and often fails to capture the precise temporal dynamics necessary for mapping CVR delay. This work addresses these gaps by combining deep learning with a novel rescaling strategy.
Methods
Deep Learning Pipeline
The researchers developed a 1D Fully Convolutional Network (FCN) trained to transform RVT signals into predicted $P_{ET}CO_2$ waveforms. Unlike models trained on resting-state data, this FCN was specifically optimized for breath-holding tasks. A specialized "peak-focused" loss function was employed to ensure the model accurately captured the hypercapnic peaks, which are the most critical segments for CVR calculation.
The "One-Trial" Calibration
To convert the arbitrary units of RVT or predicted $CO_2$ into standard mmHg, the authors introduced a rescaling method. They hypothesized that if a participant can provide just one high-quality breath-hold trial (a single clean exhalation) during a scan, that segment can serve as a "gold standard" to calibrate the entire session's predicted waveform. This significantly lowers the compliance threshold for patients.
Experimental Design
The model was trained and validated using a diverse set of breath-holding protocols to ensure generalizability. The datasets included both fixed-duration and randomized breath-hold tasks.
Results
Accuracy and Reliability
The predicted $P_{ET}CO_2$ regressor showed exceptional performance, achieving a mean Fisher’s z-transformed correlation of 1.14 with ground-truth measurements. In terms of CVR amplitude, the median error in gray matter was only 0.037 %BOLD/mmHg—a level of precision sufficient to distinguish between different age groups or clinical states.
Sensitivity to Pathology
The method's clinical utility was tested in a patient with Moyamoya disease, characterized by narrowed intracranial arteries and significant CVR delays. The predicted $CO_2$ regressor successfully identified regions of extreme vascular delay with a Dice similarity coefficient (DSC) of 0.66 compared to ground truth. In contrast, the raw RVT-based method failed to accurately map these delays, yielding a DSC of only 0.16.
Discussion
This study demonstrates that deep learning can effectively bridge the gap between robust but non-quantitative respiration metrics (RVT) and the precise $P_{ET}CO_2$ data required for clinical CVR mapping. By capturing the temporal "shape" of the $CO_2$ response more accurately than raw RVT, the FCN provides superior maps of CVR delay, which is often a more sensitive marker of disease than amplitude alone.
While the current model was trained on healthy individuals, future work will focus on expanding the training set to include patients with chronic respiratory or cardiovascular conditions. Ultimately, this "rescue" pipeline could be integrated into real-time fMRI monitoring, allowing technicians to save datasets that would otherwise be discarded due to poor sensor quality.
Summary updated