Assessment of data acquisition parameters, and analysis techniques for noise reduction in spinal cord fMRI data  R.L. Bosma, P.W. Stroman  Magnetic Resonance Imaging  Volume 32, Issue 5, Pages 473-481 (June 2014) DOI: 10.1016/j.mri.2014.01.007 Copyright © 2014 Elsevier Inc. Terms and Conditions

Fig. 1 A demonstration of the normalization procedure in one dataset. a) A midline slice of the raw data shown in the sagittal and coronal views, with the nine spine lines manually drawn on the image. b) The normalized template with the lines from the raw data spatially arranged to match template. Each node drawn on the raw data (*) is translated onto the normalized template (o) according to the relative distance along each line (for brainstem/brain lines), or the distance in millimeters along down the pons on the anterior line (spinal cord lines). c) An example of a single normalized midline slice. The image has been warped into a common space and maps onto the template. It is now centered in L/R and A/P directions and does not demonstrate any curvature. Magnetic Resonance Imaging 2014 32, 473-481DOI: (10.1016/j.mri.2014.01.007) Copyright © 2014 Elsevier Inc. Terms and Conditions

Fig. 2 The contrast-to-noise ratio (CNR) demonstrated at different echo times. The red line (o) is the CNR in the spinal cord, while the blue line (x) is the CNR in the brainstem. The black line reflects the theoretical CNR curve calculated for T2=75msec. Echo time=75msec provides the highest CNR and corresponds with the optimal spin-echo BOLD contrast in the spinal cord. Magnetic Resonance Imaging 2014 32, 473-481DOI: (10.1016/j.mri.2014.01.007) Copyright © 2014 Elsevier Inc. Terms and Conditions

Fig. 3 A representative example from 1 midline sagittal slice from 1 subject from data acquired using a HASTE sequence, with or without parallel imaging and at different repetition times. On the far left is an image collected without parallel imaging, TR=6.75seconds. The middle and far right images were acquired with parallel imaging on (GRAPPA) with a TR=4.5seconds and TR=3.6seconds, respectively. From this example you can visually discern that the image quality is degraded at the lowest TR. The lower half of the figure displays the deviation from each dataset from the T-value distribution (dashed line) that models errors that occur as a result of random noise. Values to the right of the curve indicate an increase in error rate at any given T value. Data acquired with parallel imaging with a TR=4.5seconds (green line) closely follows the students T distribution, indicating that the errors in this data arise primarily from random noise. Data acquired with a faster TR (=3.6seconds) and slower TR (=6.75seconds) deviate to the right of the curve indicating greater error rates in this data. Magnetic Resonance Imaging 2014 32, 473-481DOI: (10.1016/j.mri.2014.01.007) Copyright © 2014 Elsevier Inc. Terms and Conditions

Fig. 4 The AR(1) (left) and variance (right) heat maps for each data acquisition method. Hotter colours indicate areas with either a larger AR(1) correlation (left) or a higher variance (right). Areas of higher variance and greater occurrence of AR(1) fluctuations occur primarily near the spinal cord-CSF boundary. The use of parallel imaging and a faster TR reduce both components. Magnetic Resonance Imaging 2014 32, 473-481DOI: (10.1016/j.mri.2014.01.007) Copyright © 2014 Elsevier Inc. Terms and Conditions

Fig. 5 The autocorrelations and frequency spectrum averaged across all voxels in C4-C6 from one representative participant. Magnetic Resonance Imaging 2014 32, 473-481DOI: (10.1016/j.mri.2014.01.007) Copyright © 2014 Elsevier Inc. Terms and Conditions

Fig. 6 The sum of the number of active voxels of the null runs and task runs (thermal pain) conducted with (TR=4.5seconds) and without (TR=6.75seconds) parallel imaging. Active voxels were detected by means of a GLM analysis using the heat pain stimulation paradigm. The number of false positives was calculated for different pre-processing procedures; no pre-processing (none), with co-alignment (c), with co-alignment and Gibb’s ringing artifact removal (ac), with the data co-aligned and the first two principle components included in the GLM (PC2c) and with the data co-aligned and with temporal filtering (fc). The false positive voxel count extended beyond the axis of the graph when we included temporal filtering. The false positive rate was greatly decreased when no temporal filtering was applied. Both the use of parallel imaging and no parallel imaging produced similar results with respect to the number of false positive activations and the pre-processing steps had a similar influence on the two methods. Magnetic Resonance Imaging 2014 32, 473-481DOI: (10.1016/j.mri.2014.01.007) Copyright © 2014 Elsevier Inc. Terms and Conditions

Fig. 7 We assessed how each noise reduction step influenced the beta values and standard error of the mean. Each step is compared to the beta value and SEM without preprocessing (with/without) and the magnitude and direction of the influence is depicted. The arrows are heat coded such that red indicates that the step had a large influence on the measure, while yellow indicates a small influence. Magnetic Resonance Imaging 2014 32, 473-481DOI: (10.1016/j.mri.2014.01.007) Copyright © 2014 Elsevier Inc. Terms and Conditions

Fig. 8 The assessment of the accuracy of the normalization procedure with and without the medical image registration toolbox. Errors are calculated by the distance in millimeters that the cross-correlation procedure had to move the grid points in order to achieve the peak correlation. In all regions, the magnitude of errors was smaller after MIRT registration was applied. Magnetic Resonance Imaging 2014 32, 473-481DOI: (10.1016/j.mri.2014.01.007) Copyright © 2014 Elsevier Inc. Terms and Conditions