Removing inter-subject technical variability in magnetic resonance imaging studies

Abstract

Magnetic resonance imaging (MRI) intensities are acquired in arbitrary units, making scans non-comparable across sites and between subjects. Intensity normalization is a first step for the improvement of comparability of the images across subjects. However, we show that unwanted inter-scan variability associated with imaging site, scanner effect, and other technical artifacts is still present after standard intensity normalization in large multi-site neuroimaging studies. We propose RAVEL (Removal of Artificial Voxel Effect by Linear regression), a tool to remove residual technical variability after intensity normalization. As proposed by SVA and RUV [Leek and Storey, 2007, 2008, Gagnon-Bartsch and Speed, 2012], two batch effect correction tools largely used in genomics, we decompose the voxel intensities of images registered to a template into a biological component and an unwanted variation component. The unwanted variation component is estimated from a control region obtained from the cerebrospinal fluid (CSF), where intensities are known to be unassociated with disease status and other clinical covariates. We perform a singular value decomposition (SVD) of the control voxels to estimate factors of unwanted variation. We then estimate the unwanted factors using linear regression for every voxel of the brain and take the residuals as the RAVEL-corrected intensities. We assess the performance of RAVEL using T1-weighted (T1-w) images from more than 900 subjects with Alzheimer's disease (AD) and mild cognitive impairment (MCI), as well as healthy controls from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. We compare RAVEL to two intensity-normalization-only methods: histogram matching and White Stripe. We show that RAVEL performs best at improving the replicability of the brain regions that are empirically found to be most associated with AD, and that these regions are significantly more present in structures impacted by AD (hippocampus, amygdala, parahippocampal gyrus, enthorinal area, and fornix stria terminals). In addition, we show that the RAVEL-corrected intensities have the best performance in distinguishing between MCI subjects and healthy subjects using the mean hippocampal intensity (AUC=67%), a marked improvement compared to results from intensity normalization alone (AUC=63% and 59% for histogram matching and White Stripe, respectively). RAVEL is promising for many other imaging modalities.

Keywords: ADNI; Alzheimer's disease; MRI; Normalization; Scan effect.

Figure A.1. CAT plots with additional methods

(a) Like Figure 4b, but distinguish between RAVEL run on intensities normalized by White Stripe (default) and RAVEL run on intensities normalized by histogram matching. (b) Like Figure 4b, but for different numbers of unwanted factors in the RAVEL model. The pink line is for RAVEL with 2 factors, and the grey lines represent RAVEL with 3 to 15 factors. We can observe that the choice of 1 or 2 factors in the RAVEL model optimizes the replication of the voxels associated with AD.

Figure A.2. CAT plots with explicit correction for technical covariates

(a) The solid lines correspond to the CAT curves described in Figure 4b. The dotted lines correspond to the CAT curves for the data corrected by each of the normalization method and corrected for the following technical covariates: field strength (1.5T or 3T) and scanner manufacturer (Siemens, GE, or Philips). The correction was made by adjusting for the technical covariates in the multiple linear model analysis framework. (b) is similar to (a), but also adjusting for scanner site.

Figure A.3. Voxel-level p-value maps from AD vs. healthy patient differential analysis

At each voxel, we computed a t-statistic for testing a difference in intensities between AD and healthy patients. For each normalization method, we report the negative log p-values from the t-test. We include at the top of the figure the template for anatomical reference.

Figure A.4. Number of false positives for each method

(a) Number of false positives (voxels that do not fall into the silver standard) for a range of values of k ≤ 10, 000.

Figure A.5

Location of the false positives for the RAVEL-corrected intensities, for the top k=10,000 voxels associated with AD.

Figure 1

Schematic showing the RAVEL pipeline. The steps shown in the blue region are standard preprocessing steps that can be run in parallel. The green region shows the RAVEL algorithm.

Figure 2. Estimation of technical variability using CSF control voxels

(a) The voxels selected in the RAVEL model as control voxels for CSF are shown in blue overlaid on the template; the control voxels were selected as voxels classified as CSF for every subject. (b) Heatmap of the RAVEL coefficient γ̂ from Equation 1 depicted on the template, using b = 1 in Equation 1. The coefficient depends on the brain tissue, with a high coefficient for voxels in CSF (yellow regions), a moderate coefficient in GM (orange and lighter red) and a low coefficient for WM (darker red).

Figure 3. Effect of RAVEL on the histograms of intensities

Rows correspond to different preprocessing steps, and columns to different brain tissues. Each curve represents the histogram of intensities for one subject.

Figure 4. RAVEL improves replicability of voxels associated with AD

(a) In template space, we depict in yellow the voxels that are replicated across all random splittings, from the list of the top 50,000 associated with AD. (b) Mean CAT curves for association with AD with 95% confidence bands. (c) Number of voxels replicated for each method in (a). RAVEL shows excellent performance at replicating the discovery of regions of the brain associated with AD.

Figure 5. The top voxels associated with AD are enriched for the hippocampus and parahippocampal regions

(a) For the top k voxels associated with AD (x-axis), the solid lines display the number of voxels out of the k voxels falling into five structures known to be associated with the progression of AD: the hippocampus, amygdala, enthorinal cortex, fornix and stria terminalis and parahippocampal gyrus. The dotted line represents the number of voxels expected by chance only. The shaded areas represent 95% confidence bands computed using 100 bootstrapped samples. (b) From the t-statistics measuring the association of the voxel intensities with AD, we present the pseudo-ROC curves for classifying a voxel as a member of the five regions described in (a). RAVEL shows significantly better sensitivity and specificity than the other methods for detecting hippocampus and parahippocampal changes associated with AD.

Figure 6. RAVEL improves the prediction of AD and MCI

(a) The mean hippocampus intensity was used to predict AD. The AUC is 81.7 % for RAVEL, 74.9% for histogram matching, 64.4% for White Stripe and 57.0% for no normalization, with 95% CIs [77.6, 85.4], [70.4, 79.2], [58.9, 69.0] and [52.1, 62.0] respectively. (b) The mean hippocampus intensity was used to predict MCI. The AUC is 67.3% for RAVEL, 63.4% for histogram matching, 59.0% for White Stripe and 52.9% for no normalization with 95% CIs [63.1, 71.3], [59.6, 67.7], [54.8, 63.4] and [48.4, 57.3] respectively.