Predicting PETCO2 from RVT in Breath-Hold fMRI for Cerebrovascular Reactivity Mapping

2610 

Abstract Submission 

Rebecca Clements¹, Kimberly Hemmerling¹, Elijah Huang¹, Max Wang¹, Molly Bright¹

¹Northwestern University, Chicago, IL

Rebecca Clements  
Northwestern University
Chicago, IL

Kimberly Hemmerling  
Northwestern University
Chicago, IL

Elijah Huang  
Northwestern University
Chicago, IL

Max Wang  
Northwestern University
Chicago, IL

Molly Bright  
Northwestern University
Chicago, IL

Cerebrovascular reactivity (CVR), the response of blood vessels to a vasoactive stimulus, provides important information about cerebrovascular health[8,9]. Breath-hold tasks during fMRI provide a robust method for measuring CVR; typically, end-tidal CO2 (PETCO2) is measured and used to calculate CVR in standard units (%BOLD/mmHg). While PETCO2 allows for quantitative comparisons across subjects and sessions, it requires additional equipment and task compliance[10]. Respiration volume per time (RVT) is an alternative to PETCO2 that circumvents these requirements[2]; however, it is recorded in arbitrary units and only allows for relative CVR comparisons between brain regions of a single scan[10]. This work provides a more feasible method for quantitative CVR mapping by using machine learning to predict PETCO2 from RVT in breath-hold task data.

185 physiological datasets were collected in 37 healthy adults; each consisted of expired CO2 pressure and chest position traces simultaneously recorded via nasal cannula, gas analyzer, and respiration belt at 100 Hz during a 5–7-minute repeated breath-hold task: 24-s paced breathing, 10-20-s breath-hold, 2-s exhale, and 6-s recovery. End-tidal CO2 peaks and belt trace minima/maxima were identified using a peak detection algorithm, manually verified, and used to calculate PETCO2 and RVT [2], which were downsampled to 10 Hz and convolved with the HRF and RRF, respectively[3,4]. Delay between PETCO2 and RVT was corrected by temporally shifting the PETCO2 trace by the delay with maximum cross-correlation with RVT. 31 datasets with RVT and PETCO2 cross-correlation<0.5 were excluded.

Imitating the methods of Agrawal et al.[1], a 1D fully convolutional network was created to predict demeaned PETCO2 from RVT. For comparison, a model was also created to predict standardized PETCO2 from RVT (note, this model would only allow for CVR mapping in relative units). 4 different architectures with 1 (FCN-1L), 2 (FCN-2L), 4 (FCN-4L), and 6 (FCN-6L) convolutional layers were tested. For each model, training was executed using 5-fold cross validation with a weighted mean squared error (MSE) loss function, Adam optimizer, 15 epochs, and a learning rate of 0.01[1]. Model performance was assessed on the test set (20% of the dataset), and the model with the lowest test set mean absolute error was used for the below analysis.

For 6 of the test set subjects, fMRI data was acquired with the physiological data (for detailed methods, see [5]). Using both the true and predicted demeaned PETCO2 regressors, CVR was mapped using phys2cvr[6,7]. Bland-Altman analysis was performed to assess agreement between average gray matter (GM) CVR values calculated with true and predicted demeaned PETCO2 regressors.

The highest r value of the demeaned PETCO2 prediction models in breath-hold data was 0.680 (FCN-6L), which out-performs a previous study that predicted standardized PETCO2 from RVT in resting-state data with r=0.512[1]. Breath-holds evoke larger physiological fluctuations and thus may allow for more robust prediction of PETCO2. The demeaned and standardized PETCO2 prediction models had comparable performance (Fig.1); FCN-2L and FCN-6L produced the lowest MAE for the demeaned and standardized PETCO2 prediction models, respectively, and were used for all subsequent analysis. For one example subject, true and predicted PETCO2 regressors and the corresponding CVR maps are shown in Fig.2A/B. Across all 6 subjects, the CVR maps in standard units calculated using the demeaned predicted and true PETCO2 traces had an average GM CVR spatial correlation of 0.95±0.02. Bland-Altman analysis indicated no significant proportional bias in average GM CVR estimation(Fig.2C).

Quantitative PETCO2 can be predicted from RVT in breath-hold data and used to map CVR. Future work will assess whether a trial-by-trial version of our approach can mitigate the effects of poor task compliance and missing PETCO2 data on CVR estimation.

Classification and Predictive Modeling ²

Cerebral Metabolism and Hemodynamics ¹

Cerebral Blood Flow

Data analysis

Design and Analysis

FUNCTIONAL MRI

Machine Learning