Image processing and Quality Control for the first 10,000 brain imaging datasets from UK Biobank

Abstract

UK Biobank is a large-scale prospective epidemiological study with all data accessible to researchers worldwide. It is currently in the process of bringing back 100,000 of the original participants for brain, heart and body MRI, carotid ultrasound and low-dose bone/fat x-ray. The brain imaging component covers 6 modalities (T1, T2 FLAIR, susceptibility weighted MRI, Resting fMRI, Task fMRI and Diffusion MRI). Raw and processed data from the first 10,000 imaged subjects has recently been released for general research access. To help convert this data into useful summary information we have developed an automated processing and QC (Quality Control) pipeline that is available for use by other researchers. In this paper we describe the pipeline in detail, following a brief overview of UK Biobank brain imaging and the acquisition protocol. We also describe several quantitative investigations carried out as part of the development of both the imaging protocol and the processing pipeline.

Keywords: Big data imaging; Epidemiological studies; Image analysis pipeline; Machine learning; Multi-modal data integration; Quality control.

Fig. 1

Flowcharts legend. All linear registrations shown in the pipeline are rigid body transformations (6 degrees of freedom) in intra-subject operations and affine transformations (12 degrees of freedom) when registering to MNI template.

Fig. 2

General flowchart for the pipeline.

Fig. 3

Example of defacing. Left: Original T1 volume (non-Biobank subject). Center: Applying defacing masks. Right: Defaced T1 volume.

Fig. 4

One volunteer was scanned for testing purposes at different positions on the scanning table, depicting the effect of gradient distortion (note the warping in the neck). Correction for gradient distortion (C and D) substantially improves alignment. i: Subject placed low in the scanner (Baseline). ii: Subject placed in the centre of the scanner. iii: Subject placed higher into the scanner. A: ii (red outline) linearly registered to i (background) (Cross correlation: 0.90). B: iii (red outline) linearly registered to i (background) (Cross correlation: 0.83). C: ii (red outline) linearly registered to i (background) after GDC (Cross correlation: 0.98). D: iii (red outline) linearly registered to i (background) after GDC (Cross correlation: 0.96). Regions highlighting the largest effect (i.e. improved alignment of the ventricle) of the GDC have been zoomed to a 3:1 scale.

Fig. 5

Flowchart for the T1 processing pipeline.

Fig. 6

Flowchart for the defacing of the T1.

Fig. 7

Flowchart for the T2 FLAIR processing pipeline.

Fig. 8

Flowchart for the defacing of the T2 FLAIR.

Fig. 9

Histogram of WM intensities for a typical subject, showing that use of the bias field estimated from the T1 image can be applied to the T2 FLAIR image with similar improvement in IQR range to the one obtained in T1. Also, the histogram of this method is not very different to the one obtained by applying FAST directly to T2 FLAIR. A: T1. B: T2 FLAIR. C: Distribution of WM intensity Inter Quartile Ranges (IQRs) using different normalization methods in 78 subjects.

Fig. 10

Flowchart for the swMRI processing pipeline.

Fig. 11

Flowchart for the fieldmap generation pipeline.

Fig. 12

Flowchart for the dMRI processing pipeline.

Fig. 13

Comparison of 14 different alignment methods of FA to MNI space. We used the same methodology as de Groot et al. (2013). For each registration method, we used its estimated warp field in autoPtx to transform 27 automatically defined tracts into standard space; as discussed in de Groot et al. (2013), judging cross-subject alignment through similarity of tracts can be considered a test of alignment success that is reasonably independent of the images and cost functions used to drive the alignments. Each box plot shows the average cross-correlation over the 27 tracts for the 4950 combinations of pairs of 100 subjects. Figs. S13–S16 the supplementary material show this same plot, at the tract level. 1: FA linearly aligned to T1 + T1 non-linearly aligned to MNI. 2: FA linearly aligned to T1 + T1’s WM non-linearly aligned to MNI’s WM. 3: FA linearly aligned to T1 + T1’s GM non-linearly aligned to MNI’s GM. 4: Corrected b = 0 linearly aligned (BBR) to T1 + T1 non-linearly aligned to MNI. 5: Corrected b = 0 linearly aligned (BBR) to T1 + T1’s WM non-linearly aligned to MNI’s WM. 6: Corrected b = 0 linearly aligned (BBR) to T1 + T1’s GM non-linearly aligned to MNI’s GM. 7: FA non-linearly aligned to T1 + T1 non-linearly aligned to MNI. 8: FA non-linearly aligned to T1 + T1’s WM non-linearly aligned to MNI’s WM. 9: FA non-linearly aligned to T1 + T1’s GM non-linearly aligned to MNI’s GM. 10: FA linearly aligned (BBR) to T1 + T1 non-linearly aligned to MNI. 11: FA linearly aligned (BBR) to T1 + T1’s WM non-linearly aligned to MNI’s WM. 12: FA linearly aligned (BBR) to T1 + T1’s GM non-linearly aligned to MNI’s GM. 13: FA non-linearly aligned to FA FMRIB58 atlas via an FA study-specific template (created by aligning all the FAs to FA FMRIB58 and then averaging). 14: FA non-linearly aligned to FA FMRIB58 atlas using high-dimensional FNIRT-based warping.

Fig. 14

Degree of correlation and similarity of tracts after reducing the number of seeds per voxel in probtrackx. A: Average over 27 tracts and 5 subjects of the correlation between 2 different runs of probtrackx; this is reduced when we reduce the number of seeds per voxel. X axis is the reduction in the number of seeds with respect to the original AutoPtx configuration. Y axis is the average correlation of the pairs of tracts. When we reach 0.1× seeds per voxel, the median correlation drops below 0.999 for some tracts. Right plot is a zoom of left plot. B: Worst instance (in terms of correlation) of probtrackx in the forceps major tract with a factor of 0.1× seeds per voxel. The results remain robust even using a reduced number of seeds per voxel. Correlation between the maps for these two runs was 0.929. Tracts were binarised for visualization using a threshold of 10% of the 99th percentile. The overlap of the 2 runs is shown in blue. The difference is shown in red. FMRIB58 FA atlas is shown in the background. C: Same tract (forceps major) from the same subject with a factor of 0.3× seeds per voxel. The results improve by increasing the number of seeds per voxel. Correlation between the maps for these two runs was 0.985. FMRIB58 FA atlas is shown in the background. The final decision was to use 0.3x.

Fig. 15

Flowchart for the fMRi processing pipeline.

Fig. 16

Flowchart for the FEAT part of the pipeline.

Fig. 17

A: Manhattan plot summarising the significance of 8 million univariate association tests between IDPs and non-brain-imaging variables in the UK Biobank database from 10,000 subjects. For each non-imaging variable (i.e., each column in the plot), only the strongest association is plotted for each class of IDP, for clarity. Plotted p-values are not corrected for multiple comparisons, but the thresholds for both false-discovery-rate and Bonferroni correction are shown as dotted lines. B: High reproducibility (r = 0.62) of these associations in the original vs. new groups of subjects; each point is a given IDP - non-brain-variable pairing. The small number of points along the y = 0 axis relate to a non-imaging measure which (as a result of investigating these points in this plot) was found to be badly drifting over time.

Fig. 18

A: Relationship between age and total volume of white matter hyperintensities. Red cross shows the subject on the right. B: WMH segmentation using BIANCA on one example dataset Age: 68.5 years. Total WMH volume: 5049 $m{m}^3$.