Improving sensitivity, specificity, and reproducibility of individual brainstem activation

Abstract

Functional imaging of the brainstem may open new avenues for clinical diagnostics. However, for reliable assessments of brainstem activation, further efforts improving signal quality are needed. Six healthy subjects performed four repeated functional magnetic resonance imaging (fMRI) sessions on different days with jaw clenching as a motor task to elicit activation in the trigeminal motor nucleus. Functional images were acquired with a 7 T MR scanner using an optimized multiband EPI sequence. Activation measures in the trigeminal nucleus and a control region were assessed using different physiological noise correction methods (aCompCor and RETROICOR-based approaches with variable numbers of regressors) combined with cerebrospinal fluid or brainstem masking. Receiver-operating characteristic analyses accounting for sensitivity and specificity, activation overlap analyses to estimate the reproducibility between sessions, and intraclass correlation analyses (ICC) for testing reliability between subjects and sessions were used to systematically compare the physiological noise correction approaches. Masking the brainstem led to increased activation in the target ROI and resulted in higher values for the area under the curve (AUC) as a combined measure for sensitivity and specificity. With the highest values for AUC, activation overlap, and ICC, the most favorable physiological noise correction method was to control for the cerebrospinal fluid time series (aCompCor with one regressor). Brainstem motor nuclei activation can be reliably identified using high-field fMRI with optimized acquisition and processing strategies—even on single-subject level. Applying specific physiological noise correction methods improves reproducibility and reliability of brainstem activation encouraging future clinical applications.

fMRI studies evaluating central respiratory control in humans

“…Denoising is particularly important in distinguishing between the true response of respiratory centers and the spill-over effect from neighboring vessels (Khalili-Mahani et al, 2013). Only few structures within the brainstem can be reliably identified on the traditional MRI (Matt et al, 2019). Parcellation techniques, widely used for segmenting MRIs, usually have one single label for the brainstem, or in some atlases, the brainstem is segmented into midbrain, pons, and medulla.…”

Unsupervised physiological noise correction of fMRI data using phase and magnitude information (PREPAIR)

“…The brainstem, for instance, is involved in disorders affecting autonomic disfunctions (Brook and Julius, 2000), affective disorders (Paul and Lowry, 2013), migraine (Denuelle and Fabre, 2013), and Parkinson’s disease (Braak et al, 2003; Holiga et al, 2015; Tison and Meissner, 2014), but is subject to high levels of physiological noise. The proximity of the brainstem to the fourth ventricle and arteries has made it the focus of effort to reduce the variance of contaminated voxels and improve sensitivity (Harvey et al, 2008; Beissner et al, 2011; Matt et al, 2019) to allow the depiction of quite small nuclei (D’Ardenne et al, 2008; Thompson et al, 2006). This can be achieved without the need for external physiological recordings using an anatomically-defined mask of brainstem and ICA (Beissner et al, 2014).…”

“…Susceptibility variations during breathing engender small changes in the magnetic field which lead to artifacts in the BOLD signal (Raj et al, 2001), as do signals related to the chest moving at the respiration frequency, which also translates into head motion (Birn et al, 2008; Krüger and Glover, 2001). Moreover, the pulsatile nature of the cardiac cycle causes variations in the blood flow as well as tissue movements which can contaminate the measured MR signal in voxels containing major blood vessels (Dagli et al, 1999), while the brainstem is particularly affected by physiological noise due to its proximity to the fourth ventricle and arteries (Harvey et al, 2008; Beissner et al, 2011; Matt et al, 2019).…”

“…Moreover, the pulsatile nature of the cardiac cycle causes variations in the blood flow as well as tissue movements which can contaminate the measured MR signal in voxels containing major blood vessels (Dagli et al, 1999), while the brainstem is particularly affected by physiological noise due to its proximity to the fourth ventricle and arteries (Harvey et al, 2008;Beissner et al, 2011;Matt et al, 2019).…”

Brainstem fMRI

“…These brainstem masks are manually delineated on the basis of mean BOLD images, whereas a CSF mask can be automatically generated to extract nuisance regressors for the aCompCor-based method. The aCompCor approach regresses signal changes in the CSF out of BOLD signals in the statistical model because these signal changes are unlikely elicited by neuronal activities [30].…”