Crop Filling: artefact removal for real‐world clinical neuroimaging data

Abstract

AbstractBackgroundThe use of standard clinical neuroimaging data for quantitative research is often precluded by imaging artefacts caused by real‐world limitations. For example, partial coverage (cropping) of T1‐weighted MRI scans can occur when a memory clinic prioritises visual ratings on T2‐weighted MRI, and so the T1 scan field‐of‐view is reduced to minimise the overall acquisition time. This study investigated cropping‐induced biases in T1‐MRI volumetric measurements and proposed and validated a solution.MethodThis study involved 1.5T T1‐weighted and T2‐weighted MRI data from 577 participants in the Alzheimer’s Disease Neuroimaging Initiative (ADNI): 173 cognitively normal (CN); 262 mild cognitive impairment (MCI); 114 probable Alzheimer’s disease dementia (AD). The ADNI data was preprocessed to resemble clinical data seen in the CODEC dataset from the Essex Memory Clinic in the UK. The T1w scans were artificially cropped to a field of view of 120mm laterally. The T2w scans were down‐sampled to have axial slice thickness of 5mm. FreeSurfer’s synthSR tool was used to synthesize a super‐resolved T1wSR from the down‐sampled T2w. The T1wSR image was registered to the cropped T1w using mri_robust_registration, with the data from the T1wSR used to fill the missing voxels in the cropped image. FreeSurfer’s recon‐all tool was used to estimate regional brain volumes in cropped , crop‐filled and original T1w images, with results compared using Bland‐Altman analysis, R2 correlation metric, and T‐test between original and cropped/crop‐filled values.ResultFigure 1 shows an example T1w from ADNI (coronal/axial view, AD patient): original (top); artificially cropped (centre); crop‐filled (bottom). Table 1 shows groupwise demographics of the ADNI participants and overall experimental results. Figure 2‐5 shows Bland‐Altman plots and R2 correlation metric for the volumes of three AD‐relevant brain regions: hippocampus, middle temporal gyrus, and ventricles. Results suggest that the proposed crop‐filling solution works well: low percentage bias (<1% crop‐filled; ∼2% cropped), high correlation (R2∼0.9 crop‐filled; ∼0.85 cropped).ConclusionThe new crop‐filling methodology was effective at removing this specific artefact seen in clinical neuroimaging data. This is an important step towards making real‐world neuroimaging data amenable to quantitative analyses.