Program Information
Denoising of Resting State MRI Signal Fluctuation Using Machine Classifiers for Cerebrovascular Reactivity Mapping
E Gates1*, A Hsu2 , P Wang3 , P Hou1 , R Colen1 , A Kumar1 , S Prabhu1 , H Liu1 , (1) The University of Texas MD Anderson Cancer Center, Houston, TX, (2) Institute of Biomedical Electronics and Bioinformatics, National Taiwan University, Taipei, Taiwan, (3) Graduate Program in Medical Physics, University of Wisconsin, Madison, WI
Presentations
TU-AB-601-7 (Tuesday, August 1, 2017) 7:30 AM - 9:30 AM Room: 601
Purpose: To assess a machine learning approach for automatic denoising of resting state fMRI data and improve the use of resting state fluctuation amplitude (RSFA) for indicating brain regions with potential cerebrovascular uncoupling in presurgical fMRI of glioma patients.
Methods: Resting state fMRI datasets underwent (in order) timing and motion correction, detrending, brain extraction, spatial smoothing (4mm FWHM), and high-pass filtering (9 mHz). The voxel-wise temporal standard deviations were calculated to generate the RSFA map following manual and automatic denosing procedures. Independent component analysis (ICA) and machine learning classification was performed using FSL (FIX v1.06, Oxford, UK). For each of five patients, components were classified by hand for ground truth. Leave-one-out cross validation and receiver operating characteristic were used to assess classifier performance. Breath-hold MRI were performed to generate cerebrovascular response (CVR) maps for comparison. After removing artifact components, the agreement of RSFA and CVR maps were assessed. RSFA maps were thresholded such that the total activated fraction of gray matter in CVR and RSFA maps were equal.
Results: The machine classifier identified true noise components with an average area under curve (AUC) of 0.83 (range 0.74 – 0.93). None of the noise components identified by the machine classifier were definitively related to neuronal activity. The Dice similarity between the RSFA and CVR maps inside gray matter and lesion ROIs after machine denoising was effectively identical to hand denoising. The similarity inside the lesion increased slightly in a majority of cases but was not statistically significant. The computed Dice similarity was also fairly insensitive to classifier aggressiveness.
Conclusion: The machine classifier performs very well in identifying artifact components in ICA of RSFA in diseased brain. Furthermore, the use of an automatic denoising process does not negatively impact the agreement of RSFA maps with CVR maps compared to manual ICA denoising.
Contact Email: