Cardiac MRI preprint: image boundaries and processing order change genetic association results
Cardiac MRI preprint: image boundaries and processing order change genetic association results
On September 4, Devansh Pandey and Vagheesh Narasimhan posted a preprint analyzing cine MRI scans from 69 932 UK Biobank participants. Their self-supervised model reconstructed masked portions of MRI videos without clinician-provided labels. They then used the model’s numerical features to search for associations with DNA variants.
Cine MRI is a sequence of images showing the heart contracting. The heart occupies only about a quarter to a third of each image; the rest includes the chest, spine, and features of image acquisition. A model trained to reconstruct the full image learns from this information along with the heart’s shape and motion.
The model converted three cine MRI views from each participant into a set of 2 304 numerical features. The researchers compressed these features into 20 principal components, representing the largest differences between participants, and tested these components for associations with DNA variants. The full feature set predicted imaging site with R² = 0,55 and ejection fraction, the proportion of blood the left ventricle pumps out with each contraction, with R² = 0,37. A higher R² indicates a more accurate prediction.
Genetic association analysis tests the relationship between DNA variants and a numerical feature defined beforehand; image boundaries and processing before the analysis determine what that feature describes. With full images, differences in body build and imaging conditions can enter the principal components alongside differences in the heart itself.
The authors cropped the images to the region containing the moving heart, removing about 65–75% of the chest field of view. They then removed the contributions of age, sex, body mass index, height, and imaging site from all 2 304 features. Only then did they derive a new set of 20 principal components for genetic association analysis.
Adjustment after compression acts on the principal components that the compression step has already selected. In this study, cropping and adjustment before compression yielded 101 genomic loci, including all 11 prespecified heart-related loci; adjustment after compression yielded 60 loci and one candidate. In a held-out sample of 13 717 participants whose scans had not been used to train the model, effect estimates for the lead variants agreed with those from the full sample.
In this analysis, image boundaries and processing order determined which differences the model encoded as features for testing associations with DNA variants.