Though in widespread clinical and research use as a tool to evaluate brain function, functional magnetic resonance imaging (FMRI) data is severely contaminated by noise, due in large part to physiologic noise caused by respiratory and cardiac variations over time. This dissertation attempts to better characterize several physiologic noise correction techniques applied to pain FMRI data. Three studies are described that collectively work toward determining an optimal physiologic noise correction algorithm to be used in future pain FMRI studies.
First, a novel algorithm, RetroSLICE, is described that uses linear regression to correct acquired images for signal intensity fluctuations correlated to measured respiratory, cardiac, and capnometry variations. The impact of this technique was assessed for a 1.5 T pain FMRI experiment. Each physiologic noise regressor used as a part of the RetroSLICE algorithm independently resulted in a decrease in timecourse variance and an improvement in model fit. Combined correction for the instantaneous effects of respiratory and cardiac variations caused a 5.4% decrease in signal variance and increased model fit (mean R2a) by 65%. The addition of ETCO2 correction as part of the general linear model led to 39% further improvement in model fit. Each of these corrections also caused changes in the group activation map.
Next, an optimal transfer function between ETCO2 level and BOLD signal changes was empirically determined using FMRI data in which paced breathing forced a 35% decrease in ETCO2. ETCO2 data convolved with this optimized response function was compared to another measure, the respiratory volume over time (RVT) convolved with an optimized respiration response function. When regressed against FMRI data collected during a breathing modulation task, ETCO2 was more strongly and diffusely correlated to the data than RVT. Conversely, when the same comparative analysis was performed on pain FMRI data, RVT was more strongly correlated than ETCO2. In both cases, allowing ± 2 s flexibility in the response function peak times did not change the relative correlation to the MR data of the ETCO2 compared to the RVT timecourses.
Finally, the well-known physiologic noise correction algorithm, RETROICOR, was implemented on pain FMRI data collected at 1.5 and 3.0 T. Respiratory and cardiac correction with Fourier series phase fitting caused an 8.2% decrease in signal variance and a 227% increase in model fit at 1.5 T, indicating performance superior to RetroSLICE. At 3.0 T, significantly greater improvements were seen: a 10.4% decrease in signal noise and 240% increase in mean R2a. ETCO2 correction applied with the optimized response function previously determined caused insignificant changes in noise reduction and model fit. Further exploration of the properties of the RETROICOR algorithm showed no difference in impact when applied with physiologic input data sampled at a much higher rate or when accounting for the interleaved slice acquisition order. These findings suggest that RETROICOR should be included as a part of the physiologic noise correction procedure in pain FMRI studies at 1.5 and 3.0 T.