Reconstruction of respiratory variation signals from fMRI data

Physiological Confounds in BOLD-fMRI and Their Correction

Functional magnetic resonance imaging (fMRI) has opened new frontiers in neuroscience by instrumentally driving our understanding of brain function and development. Despite its substantial successes, fMRI studies persistently encounter obstacles stemming from inherent, unavoidable physiological confounds. The adverse effects of these confounds are especially noticeable with higher magnetic fields, which have been gaining momentum in fMRI experiments. This review focuses on the four major physiological confounds impacting fMRI studies: low-frequency fluctuations in both breathing depth and rate, low-frequency fluctuations in the heart rate, thoracic movements, and cardiac pulsatility. Over the past three decades, numerous correction techniques have emerged to address these challenges. Correction methods have effectively enhanced the detection of task-activated voxels and minimized the occurrence of false positives and false negatives in functional connectivity studies. While confound correction methods have merit, they also have certain limitations. For instance, model-based approaches require externally recorded physiological data that is often unavailable in fMRI studies. Methods reliant on independent component analysis, on the other hand, need prior knowledge about the number of components. Machine learning techniques, although showing potential, are still in the early stages of development and require additional validation. This article reviews the mechanics of physiological confound correction methods, scrutinizes their performance and limitations, and discusses their impact on fMRI studies.

Machine learning-based estimation of respiratory fluctuations in a healthy adult population using resting state BOLD fMRI and head motion parameters

PURPOSE

External physiological monitoring is the primary approach to measure and remove effects of low-frequency respiratory variation from BOLD-fMRI signals. However, the acquisition of clean external respiratory data during fMRI is not always possible, so recent research has proposed using machine learning to directly estimate respiratory variation (RV), potentially obviating the need for external monitoring. In this study, we propose an extended method for reconstructing RV waveforms directly from resting state BOLD-fMRI data in healthy adult participants with the inclusion of both BOLD signals and derived head motion parameters.

METHODS

In the proposed method, 1D convolutional neural networks (1D-CNNs) used BOLD signals and head motion parameters to reconstruct the RV waveform for the whole fMRI scan time. Resting-state fMRI data and associated respiratory records from the Human Connectome Project in Young Adults (HCP-YA) dataset are used to train and test the proposed method.

RESULTS

Compared to using only BOLD-fMRI data for a CNN input, this approach yielded improvements of 14% in mean absolute error, 24% in mean square error, 14% in correlation, and 12% in dynamic time warping. When tested on independent datasets, the method demonstrated generalizability, even in data with different TRs and physiological conditions.

CONCLUSION

This study shows that the respiratory variations could be reconstructed from BOLD-fMRI data in the young adult population, and its accuracy could be improved using supportive data such as head motion parameters. The method also performed well on independent datasets with different experimental conditions.

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

Cerebrovascular disease is a leading cause of death globally. Prevention and early intervention are known to be the most effective forms of its management. Non-invasive imaging methods hold great promises for early stratification, but at present lack the sensitivity for personalized prognosis. Resting-state functional magnetic resonance imaging (rs-fMRI), a powerful tool previously used for mapping neural activity, is available in most hospitals. Here we show that rs-fMRI can be used to map cerebral hemodynamic function and delineate impairment. By exploiting time variations in breathing pattern during rs-fMRI, deep learning enables reproducible mapping of cerebrovascular reactivity (CVR) and bolus arrival time (BAT) of the human brain using resting-state CO2 fluctuations as a natural "contrast media". The deep-learning network is trained with CVR and BAT maps obtained with a reference method of CO2-inhalation MRI, which includes data from young and older healthy subjects and patients with Moyamoya disease and brain tumors. We demonstrate the performance of deep-learning cerebrovascular mapping in the detection of vascular abnormalities, evaluation of revascularization effects, and vascular alterations in normal aging. In addition, cerebrovascular maps obtained with the proposed method exhibit excellent reproducibility in both healthy volunteers and stroke patients. Deep-learning resting-state vascular imaging has the potential to become a useful tool in clinical cerebrovascular imaging.

Less is more: balancing noise reduction and data retention in fMRI with data-driven scrubbing

Functional MRI (fMRI) data may be contaminated by artifacts arising from a myriad of sources, including subject head motion, respiration, heartbeat, scanner drift, and thermal noise. These artifacts cause deviations from common distributional assumptions, introduce spatial and temporal outliers, and reduce the signal-to-noise ratio of the data-all of which can have negative consequences for the accuracy and power of downstream statistical analysis. Scrubbing is a technique for excluding fMRI volumes thought to be contaminated by artifacts and generally comes in two flavors. Motion scrubbing based on subject head motion-derived measures is popular but suffers from a number of drawbacks, among them the need to choose a threshold, a lack of generalizability to multiband acquisitions, and high rates of censoring of individual volumes and entire subjects. Alternatively, data-driven scrubbing methods like DVARS are based on observed noise in the processed fMRI timeseries and may avoid some of these issues. Here we propose "projection scrubbing", a novel data-driven scrubbing method based on a statistical outlier detection framework and strategic dimension reduction, including independent component analysis (ICA), to isolate artifactual variation. We undertake a comprehensive comparison of motion scrubbing with data-driven projection scrubbing and DVARS. We argue that an appropriate metric for the success of scrubbing is maximal data retention subject to reasonable performance on typical benchmarks such as the validity, reliability, and identifiability of functional connectivity. We find that stringent motion scrubbing yields worsened validity, worsened reliability, and produced small improvements to fingerprinting. Meanwhile, data-driven scrubbing methods tend to yield greater improvements to fingerprinting while not generally worsening validity or reliability. Importantly, however, data-driven scrubbing excludes a fraction of the number of volumes or entire sessions compared to motion scrubbing. The ability of data-driven fMRI scrubbing to improve data retention without negatively impacting the quality of downstream analysis has major implications for sample sizes in population neuroscience research.

Direct machine learning reconstruction of respiratory variation waveforms from resting state fMRI data in a pediatric population

In many functional magnetic resonance imaging (fMRI) studies, respiratory signals are unavailable or do not have acceptable quality due to issues with subject compliance, equipment failure or signal error. In large databases, such as the Human Connectome Projects, over half of the respiratory recordings may be unusable. As a result, the direct removal of low frequency respiratory variations from the blood oxygen level-dependent (BOLD) signal time series is not possible. This study proposes a deep learning-based method for reconstruction of respiratory variation (RV) waveforms directly from BOLD fMRI data in pediatric participants (aged 5 to 21 years old), and does not require any respiratory measurement device. To do this, the Lifespan Human Connectome Project in Development (HCP-D) dataset, which includes respiratory measurements, was used to both train a convolutional neural network (CNN) and evaluate its performance. Results show that a CNN can capture informative features from the BOLD signal time course and reconstruct accurate RV timeseries, especially when the subject has a prominent respiratory event. This work advances the use of direct estimation of physiological parameters from fMRI, which will eventually lead to reduced complexity and decrease the burden on participants because they may not be required to wear a respiratory bellows.

Comparing end-tidal CO2, respiration volume per time (RVT), and average gray matter signal for mapping cerebrovascular reactivity amplitude and delay with breath-hold task BOLD fMRI

Cerebrovascular reactivity (CVR), defined as the cerebral blood flow response to a vasoactive stimulus, is an imaging biomarker with demonstrated utility in a range of diseases and in typical development and aging processes. A robust and widely implemented method to map CVR involves using a breath-hold task during a BOLD fMRI scan. Recording end-tidal CO₂ (P_ETCO₂) changes during the breath-hold task is recommended to be used as a reference signal for modeling CVR amplitude in standard units (%BOLD/mmHg) and CVR delay in seconds. However, obtaining reliable P_ETCO₂ recordings requires equipment and task compliance that may not be achievable in all settings. To address this challenge, we investigated two alternative reference signals to map CVR amplitude and delay in a lagged general linear model (lagged-GLM) framework: respiration volume per time (RVT) and average gray matter BOLD response (GM-BOLD). In 8 healthy adults with multiple scan sessions, we compare spatial agreement of CVR maps from RVT and GM-BOLD to those generated with P_ETCO₂. We define a threshold to determine whether a P_ETCO₂ recording has "sufficient" quality for CVR mapping and perform these comparisons in 16 datasets with sufficient P_ETCO₂ and 6 datasets with insufficient P_ETCO₂. When P_ETCO₂ quality is sufficient, both RVT and GM-BOLD produce CVR amplitude maps that are nearly identical to those from P_ETCO₂ (after accounting for differences in scale), with the caveat they are not in standard units to facilitate between-group comparisons. CVR delays are comparable to P_ETCO₂ with an RVT regressor but may be underestimated with the average GM-BOLD regressor. Importantly, when P_ETCO₂ quality is insufficient, RVT and GM-BOLD CVR recover reasonable CVR amplitude and delay maps, provided the participant attempted the breath-hold task. Therefore, our framework offers a solution for achieving high quality CVR maps in both retrospective and prospective studies where sufficient P_ETCO₂ recordings are not available and especially in populations where obtaining reliable measurements is a known challenge (e.g., children). Our results have the potential to improve the accessibility of CVR mapping and to increase the prevalence of this promising metric of vascular health.

An imaging-based method of mapping multi-echo BOLD intracranial pulsatility

PURPOSE

Cardiac-related intracranial pulsatility may relate to cerebrovascular health, and this information is contained in BOLD MRI data. There is broad interest in methods to isolate BOLD pulsatility, and the current study examines a deep learning approach.

METHODS

Multi-echo BOLD images, respiratory, and cardiac recordings were measured in 55 adults. Ground truth BOLD pulsatility maps were calculated with an established method. BOLD fast Fourier transform magnitude images were used as temporal-frequency image inputs to a U-Net deep learning model. Model performance was evaluated by mean squared error (MSE), mean absolute error (MAE), structural similarity index (SSIM), and mutual information (MI). Experiments evaluated the influence of input channel size, an age group effect during training, dependence on TE, performance without the U-Net architecture, and importance of respiratory preprocessing.

RESULTS

The U-Net model generated BOLD pulsatility maps with lower MSE as additional fast Fourier transform input images were used. There was no age group effect for MSE (P > 0.14). MAE and SSIM metrics did not vary across TE (P > 0.36), whereas MI showed a significant TE dependence (P < 0.05). The U-Net versus no U-Net comparison showed no significant difference for MAE (P = 0.059); however, SSIM and MI were significantly different between models (P < 0.001). Within the insula, the cross-correlation values were high (r > 0.90) when comparing the U-Net model trained with/without respiratory preprocessing.

CONCLUSION

Multi-echo BOLD pulsatility maps were synthesized from a U-net model that was trained to use temporal-frequency BOLD image inputs. This work adds to the deep learning methods that characterize BOLD physiological signals.

Generating dynamic carbon-dioxide traces from respiration-belt recordings: Feasibility using neural networks and application in functional magnetic resonance imaging

INTRODUCTION

In the context of functional magnetic resonance imaging (fMRI), carbon dioxide (CO2) is a well-known vasodilator that has been widely used to monitor and interrogate vascular physiology. Moreover, spontaneous fluctuations in end-tidal carbon dioxide (PETCO2) reflects changes in arterial CO2 and has been demonstrated as the largest physiological noise source for denoising the low-frequency range of the resting-state fMRI (rs-fMRI) signal. However, the majority of rs-fMRI studies do not involve CO2 recordings, and most often only heart rate and respiration are recorded. While the intrinsic link between these latter metrics and CO2 led to suggested possible analytical models, they have not been widely applied.

METHODS

In this proof-of-concept study, we propose a deep-learning (DL) approach to reconstruct CO2 and PETCO2 data from respiration waveforms in the resting state.

RESULTS

We demonstrate that the one-to-one mapping between respiration and CO2 recordings can be well predicted using fully convolutional networks (FCNs), achieving a Pearson correlation coefficient (r) of 0.946 ± 0.056 with the ground truth CO2. Moreover, dynamic PETCO2 can be successfully derived from the predicted CO2, achieving r of 0.512 ± 0.269 with the ground truth. Importantly, the FCN-based methods outperform previously proposed analytical methods. In addition, we provide guidelines for quality assurance of respiration recordings for the purposes of CO2 prediction.

DISCUSSION

Our results demonstrate that dynamic CO2 can be obtained from respiration-volume using neural networks, complementing the still few reports in DL of physiological fMRI signals, and paving the way for further research in DL based bio-signal processing.

A comprehensive investigation of physiologic noise modeling in resting state fMRI; time shifted cardiac noise in EPI and its removal without external physiologic signal measures

Hemodynamic cardiac and respiratory-cycle fluctuations are a source of unwanted non-neuronal signal components, often called physiologic noise, in resting state (rs-) fMRI studies. Here, we use image-based retrospective correction of physiological motion (RETROICOR) with externally measured physiologic signals to investigate cardiac and respiratory hemodynamic phase functions reflected in rs-fMRI data. We find that the cardiac phase function is time shifted locally, while the respiratory phase function is described as single, fixed phase form across the brain. In light of these findings, we propose an update to Physiologic EStimation by Temporal ICA (PESTICA), our publically available software package that estimates physiologic signals when external physiologic measures are not available. This update incorporates: 1) auto-selection of slicewise physiologic regressors and generation of physiologic fixed phase regressors with total slices/TR sampling rate, 2) Fourier series expansion of the cardiac fixed phase regressor to account for time delayed cardiac noise 3) removal of cardiac and respiratory noise in imaging data. We compare the efficacy of the updated method to RETROICOR.

Structure can predict function in the human brain: a graph neural network deep learning model of functional connectivity and centrality based on structural connectivity

Although functional connectivity and associated graph theory measures (e.g., centrality; how centrally important to the network a region is) are widely used in brain research, the full extent to which these functional measures are related to the underlying structural connectivity is not yet fully understood. Graph neural network deep learning methods have not yet been applied for this purpose, and offer an ideal model architecture for working with connectivity data given their ability to capture and maintain inherent network structure. Here, we applied this model to predict functional connectivity from structural connectivity in a sample of 998 participants from the Human Connectome Project. Our results showed that the graph neural network accounted for 89% of the variance in mean functional connectivity, 56% of the variance in individual-level functional connectivity, 99% of the variance in mean functional centrality, and 81% of the variance in individual-level functional centrality. These results represent an important finding that functional centrality can be robustly predicted from structural connectivity. Regions of particular importance to the model's performance as determined through lesioning are discussed, whereby regions with higher centrality have a higher impact on model performance. Future research on models of patient, demographic, or behavioural data can also benefit from this graph neural network method as it is ideally-suited for depicting connectivity and centrality in brain networks. These results have set a new benchmark for prediction of functional connectivity from structural connectivity, and models like this may ultimately lead to a way to predict functional connectivity in individuals who are unable to do fMRI tasks (e.g., non-responsive patients).