Machine Learning For High-Throughput Quality Control Of Cardiac Imaging

Pixabay License | Source: Michal Jarmoluk , no changes made.
Advertisement img

From 2014 onwards, 100,000 volunteers from the UK biobank (UKBB) were enrolled for multi-modal imaging, including magnetic resonance (MRI) of the brain, the heart and the full body. More than 20,000 participants have been MRI scanned for the heart so far. The dataset is intended as a reference standard and therefore quality control is essential.

In the usual patient by patient basis of hospitals quality control (QC) is performed by the MRI operator with the standard being strongly subjective. The UKBB study required a more high-throughput and standardized approach.

A fully-automated computational QC pipeline was recently developed using hybrid random forests for cardiac MRI that estimates heart coverage, inter-slice motion, and image contrast in the cardiac region. Giacomo Tarroni of Imperial College London and colleagues have now applied the QC process to the first batch of nearly 20,000 Cardiac MRI scans from the UKBB, the results of which they published in the journal Scientific Reports.

Featured Partners

The majority of image stacks covered the whole left ventricle or more; 86% with complete coverage. The scans acquired in the first year of collection had lower coverage in comparison to the subsequent years. There was no relationship between the image coverage and the subject’s weight or with body surface area.

Weight and body area was correlated with average misalignment and image stacks acquired in subjects with lower weight were less likely to have sub-optimal alignment. Self-reported cardiovascular and respiratory diseases seemed to be associated with misalignment.

There was very little variation in image contrast.

“The results show that while quality metrics are generally high throughout the whole UKBB dataset, some small differences in quality were associated with a few factors (e.g. acquisition site and date for heart coverage, and weight and body surface area of the subject being scanned for inter-slice motion). These results could be beneficial both to the scientists involved in data acquisition for large population studies like the UKBB as well as for those who use this valuable dataset for research purposes,” concluded the authors.