TL;DR
This paper tackles the challenge of accurately measuring cerebrovascular reactivity (CVR) using fMRI when participants struggle with precise breath-holding tasks, leading to poor end-tidal CO₂ (PETCO₂) data. They introduce a method to predict PETCO₂ from more easily acquired respiration data using deep learning, which, when rescaled, yields more accurate CVR maps in standard units, crucial for clinical applications.
What Problem Does This Paper Address?
Cerebrovascular reactivity (CVR) is a vital indicator of brain health, reflecting how blood vessels respond to changes in CO₂. A common way to measure CVR is with fMRI during a breath-hold task, which artificially raises blood CO₂ levels. To quantify CVR accurately in standard units (%BOLD/mmHg), researchers typically measure end-tidal CO₂ (PETCO₂) during the scan. However, this requires participants to exhale precisely before and after each hold and breathe only through their nose, a challenging task, especially for patients or children. This poor compliance leads to unreliable PETCO₂ data, limiting CVR's clinical use. While respiration volume per time (RVT) from chest belts is easier to obtain, it's in arbitrary units, preventing direct comparison across individuals. This paper aims to bridge this gap by developing methods to derive accurate CVR measurements in standard units even when PETCO₂ data quality is low.
Methods
The core of this work lies in two main strategies:
- Rescaling RVT: The researchers investigated if respiration volume per time (RVT) data, collected using a respiration belt, could be scaled to mmHg units. They used a subset of high-quality PETCO₂ data from breath-hold trials to establish a scaling factor, effectively converting RVT into a proxy for CO₂ changes.
- Deep Learning for PETCO₂ Prediction: They employed a 1D convolutional neural network (FCN) to predict PETCO₂ time series from RVT data. This model was trained on a dataset of simultaneous RVT and PETCO₂ recordings during breath-hold tasks. Crucially, they then rescaled these predicted PETCO₂ time series to mmHg using the same method as for RVT.
Both rescaled RVT and rescaled predicted PETCO₂ were then used as regressors in a standard fMRI analysis pipeline to generate CVR amplitude and delay maps.
Datasets
- In-house Training Dataset: A custom dataset comprising 245 simultaneous exhaled CO₂ (PETCO₂) and respiration effort (RVT) time series from 56 healthy participants (26 ± 4 years) collected during various breath-hold tasks. This dataset was used to train the deep learning model for PETCO₂ prediction.
- EuskalIBUR Testing Dataset: A publicly available dataset containing fMRI, PETCO₂, and RVT data from 10 participants (32 ± 6 years) across 99 scan sessions. This dataset was used to evaluate the performance of the rescaled RVT and predicted PETCO₂ regressors for CVR mapping.
Key Insights and Ideas
- Breath-hold compliance is a major bottleneck: The study highlights that even in healthy individuals, achieving consistently high-quality PETCO₂ data during breath-hold tasks is difficult, with a significant portion of trials being low-quality. This underscores the need for alternative methods.
- Deep learning can bridge the gap: The FCN model demonstrated a strong ability to predict PETCO₂ from RVT, significantly outperforming previous attempts that used resting-state data. This suggests that the larger CO₂ fluctuations during breath-holds are more informative for prediction.
- Rescaling is key for clinical relevance: Simply predicting PETCO₂ isn't enough; the ability to rescale these predictions (or RVT) into standard mmHg units is critical for enabling direct comparisons and establishing normative CVR values.
- Predicted PETCO₂ offers superior accuracy: The rescaled predicted PETCO₂ regressor consistently outperformed rescaled RVT in generating CVR amplitude and delay maps, showing higher correlations and lower errors compared to ground-truth PETCO₂ data.
- CVR delay is more sensitive to pathology: In a case study of Moyamoya disease, the predicted PETCO₂ method was better at identifying regions with increased CVR delay, suggesting its potential for detecting subtle vascular impairments.
Key Results
- The FCN model achieved a mean Pearson correlation of 0.787 ± 0.116 between predicted and measured PETCO₂ on the EuskalIBUR dataset, a significant improvement over previous resting-state prediction methods.
- Rescaled predicted PETCO₂ showed significantly higher correlation (Fisher's Z = 1.24 ± 0.36) with measured PETCO₂ than rescaled RVT (Fisher's Z = 1.14 ± 0.32) in high-quality PETCO₂ datasets.
- Group-level MAE maps indicated that rescaled predicted PETCO₂ resulted in lower errors for both CVR amplitude (median MAE in GM = 0.037 %BOLD/mmHg with 1 breath hold) and delay (median MAE in GM = 0.97 seconds) compared to rescaled RVT.
- In a Moyamoya disease case study, the predicted PETCO₂ delay map showed a Dice similarity coefficient of 0.66 with the ground-truth delay map, significantly higher than that of the RVT delay map (0.16).
Limitations and Future Work
- Rescaling variability: While using 1-3 high-quality breath holds for rescaling generally improved accuracy, a small cluster of scans showed higher errors when using only one breath hold, suggesting that the representativeness of that single breath hold's CO₂ change is important. The authors recommend using all available high-quality breath holds for rescaling if possible.
- Generalizability of the prediction model: While the model showed good generalizability, the training dataset was relatively small and primarily consisted of younger participants. Future work should include data from a wider age range and diverse breathing styles.
- Clinical validation: More extensive validation is needed to confirm the sensitivity of the proposed methods in identifying meaningful differences in CVR across various patient populations, especially those with low breath-hold task compliance.
- Model complexity: Future work could explore more complex models, such as U-Nets, or time-series forecasting approaches to potentially improve prediction accuracy and eliminate the need for a separate rescaling step.
- HRF variability: The authors acknowledge that the canonical HRF may not be universally optimal and suggest that future models could allow for more flexible HRF selection.
Summary updated