Introduction

Data sharing in neuroimaging research communities has become increasingly popular, with multiple collaborative efforts for pooling data to form large and diverse samples^1,2. The limited standardization of image acquisition protocols across sites with different scanner types and the heterogeneity of exported data structures and processing pipelines represent major challenges for Research Data Management (RDM) in multisite cohort studies³. Additionally, no clear procedure for MRI quality control (QC) in large cohort studies exists⁴. Here, we describe the methodological advances and implementation details of the MRI data processing and QC procedures developed for the European Prevention of Alzheimer Dementia (EPAD) neuroimaging dataset⁵.

Method

The EPAD v1500.0 MRI acquisition protocol
The EPAD v1500.0 release is the baseline data release of the EPAD LCS study and includes data from the first 1500 participants consented in the study⁶. The imaging protocol is composed of core and advanced sequences. The core image acquisitions (3D T1w, 3D FLAIR, 2D T2w, 2D T2star) are conducted at all sites and provide structural information. The advanced MRI protocol (resting-state fMRI, diffusion tensor imaging, arterial spin labeling, susceptibility weighted imaging) is performed in a subset of EPAD sites with sufficient expertise and equipment, and investigates brain function and structure in greater detail.
The EPAD Image Processing Pipeline
The EPAD image processing pipeline consisted of four modules (Figure 1): 1) Curation of raw DICOM files, including conversion to Neuroimaging Informatics Technology Initiative (NifTI) format; 2) Image pre-processing of core and advanced sequences; 3) Semi-automatic quality control (QC) 4) Image-derived Phenotypes (IDPs) computation⁷, i.e. extraction of summary measures from images.
DICOM data were automatically parsed for series description identification using regular expressions and sorted into folders with a unique structure throughout the dataset. Pre-processing of core and advanced sequences is shown in Figure 2, and was implemented in ExploreASL 1.0.2⁸, with the addition of BaMos⁹, FSL 6¹⁰ and SPM 12¹¹.
We developed and integrated ExploreQC: a two-module toolbox for the implementation of semi-automated MRI Quality Control of 3D T1w scans. The parameter estimation module was used to compute QC parameters defined over five Image Features Domains (IFDs): Noise, Motion, Inhomogeneity, Asymmetry, and Descriptives. The visualization module creates an interactive dashboard reporting within and between site parameter distributions and flags the 15% of scans showing the highest deviation from the site mean as “needing visual QC”. The flagged, and an equivalent number of non-flagged, images were visually inspected and judged as “good”, “moderate”, or “poor” quality, by a blinded rater. As an explorative step, we built a multiple logistic regression model with QC parameters predicting the output of the visual judgement and performed a stepwise backward feature selection based on Akaike Information Criteria to investigate the informative character of each metric.
Core sequences’ IDPs include regional volume and thickness values using the FreeSurfer¹² and LEAP¹³ segmentation pipelines, and global white matter hyperintensity (WMH) load as measured by BaMoS⁹. A group-level independent component analysis (ICA) was performed with FSL melodic on the processed functional MRI (fMRI) dataset to extract mean functional connectivity for each resting-state network (Figure 3). FSL tract-based spatial statistics (TBSS) pipeline was used on Diffusion Tensor Imaging (DTI) derived fractional anisotropy (FA) images to compute 48 regional features according to the JHU ICBM-DTI-81 atlas¹⁴ (Figure 4). Mean cerebral blood flow (CBF) and spatial coefficient-of-variation (sCoV) were computed from Arterial Spin Labeling (ASL) scans as described previously⁸. To evaluate IDPs biological relevance we assessed their relationship with non-imaging metadata.

Results

Of the 1500 screened participants, 142 did not fulfill eligibility criteria and were therefore excluded from the release⁶. The resulting dataset included 1358 core sequences, 756 SWI, 842 fMRI, 831 DTI, and 858 ASL scans.
Of the 197 automatically flagged scans, 16 were categorized as “poor quality” upon visual inspection, while 51 were labelled as “moderate quality”. Of the equivalent number of non-flagged scans only one was labelled “poor” while 10 were judged “moderate”. Backward parameter elimination included 8 out of the 12 structural QC parameters in the final selected model. Five of those were significantly associated with the visual QC output (Table 1).
A total of 358 IDPs were computed from core sequences as described above. From advanced sequences we derived 119 IDPs. Figure 3 and 4 show example associations with other non-imaging metadata.

Conclusion

Here, we presented the semi-automated pipeline used to process and extract relevant features from a multimodal and multicenter MRI cohort. We introduced a QC toolbox for the computation and visualization of informative parameters for guiding image inclusion in large studies. Our work may benefit both imaging and non-imaging researchers accessing the EPAD data as well as investigators aiming to apply a similar procedure in other multi-center cohorts.

Acknowledgements

No acknowledgement found.