You are viewing the SNL 2017 Archive Website. For the latest information, see the Current Website.

 
Poster A52, Wednesday, November 8, 10:30 – 11:45 am, Harborview and Loch Raven Ballrooms

ICA-based classifiers mitigate task correlated motion artifacts for overt-speech fMRI paradigms in aphasia

Venkatagiri Krishnamurthy1,2, Lisa Krishnamurthy2,3, Kaundinya Gopinath4, Michelle Benjamin5,6, Bruce Crosson1,2,5,7;1Dept. of Neurology, Emory University, Atlanta, GA, United States, 2Center for Visual and Neurocognitive Rehabilitation, Atlanta VAMC, Decatur, GA, United States, 3Dept. of Physics & Astronomy, Georgia State University, Atlanta, GA, United States, 4Dept. of Radiology & Imaging Sciences, Emory University, Atlanta, GA, United States, 5University of Florida, Gainesville, FL, United States, 6Brooks Rehabilitation, Jacksonville, FL, United States, 7Dept. of Psychology, Georgia State University, Atlanta, GA, United States

Overt-speech fMRI paradigms are very useful in aphasia research, but are also plagued by task-correlated motion (TCM). Several studies have developed methodologies to overcome TCM artifacts, but a relatively easy and computationally time-efficient technique that maintains the specificity-sensitivity balance in detection of task-related functional activity is still lacking. The goal of this study is to develop a novel ICA-based trained classifier that is optimized to capture spatio-temporally varying TCM, whose use can be cost-effective on large datasets.Four monolingual English speaking patients with aphasia (post-stroke >6 months) were recruited. We acquired high-resolution MPRAGE structural images and six task-fMRI runs. During the task-fMRI runs, patients heard and read a semantic category and attempted to overtly generate an exemplar of that category. Functional images were corrected for slice timing, global head motion, and decomposed into temporal and spatial components using FSL MELODIC. Deconvolution on the IC components’ time-series was used to obtain an impulse response function (IRF) that were visually inspected for TCM characteristics. TCM-like IRFs were further validated for its signature by using the designed task-specific periodicity and power spectral density at task frequency (0.03Hz). In parallel, a 3D deconvolution was applied on all fMRI runs to obtain a voxel IRF, which was further correlated with the TCM-like IRF. The result was masked by a ‘TCM’ mask that included major TCM-prone brain areas, and the correlation was thresholded at 0.7. Based on above criteria, the IC component was labeled as ‘TCM-noise’ component, and this step was carried out on 10 different data sets to obtain a trained-classifier, which was applied to the remaining datasets. The data sets were processed in 4 ways: (i) no application of any TCM correction (no ICA), (ii) novel TCM classifier (TCM), (iii) standard package, i.e., AROMA (AROMA), and (iv) TCM classifier followed by AROMA (TCM+AROMA). Finally, each denoised dataset was spatially smoothed excluding CSF, and deconvolved with the task stimuli to generate a HRF and statistical parametric activation map thresholded at R2 =0.16.TCM classifier out performs the other methods both in terms of specificity and sensitivity. In the process of cleaning-up TCM, the AROMA and TCM+AROMA methods removed BOLD signal from important language areas such as left Heschl’s and angular gyrus, and right Superior Temporal gyrus, whereas the novel TCM classifier retained sensitivity in those areas. The TCM classifier also improved specificity by removing false positive activations from surrounding larger unrelated brain areas. We also noted that smoothing within grey and white matter (i.e. excluding CSF) increased specificity (i.e >R2).Our preliminary results show that TCM-specific ICA-based classifier is promising as evidenced by the improved sensitivity and specificity. Since speech-related motion has unique signatures, stock packages such as AROMA is not ideal for denoising TCM. Further, the advantage of semi-automated ICA classifier is that it requires a one-time front-end effort to hand label and train the classifiers thus reducing the burden of excessive computational and labor time to denoise each dataset. Future work will include extending this methodology to remove trial-by-trial variations in TCM.

Topic Area: Methods

Back to Poster Schedule