Synthetic Balancing of Cardiac MRI Datasets

Carles Garcia-Cabrera, Eric Arazo Sánchez, Enric Moreu, Kathleen M. Curran, Noel E. O’Connor, Kevin McGuinness

Abstract: Possessing a balanced dataset to train the segmentation models used in clinical products is hugely important to enhance the performance in any given subset of patients. Due to demographics and uneven distribution of conditions of the subjects, this is not usually the case. In this work, we propose a novel method that synthetically balances a training dataset by applying deformations to an atlas. In particular, we focus on modelling two diseases that have distinct heart morphologies. Once the atlas is processed, we obtain different slice cuts from it and apply style transfer to make it appear as a real short-axis MRI scan. We then add those synthetic scans to our training set for the segmentation network. Our experiment compares the performance of the distinct trained models in a test set. Additionally, we also present the results on subsets of the test set representing the modelled pathologies. We found that using synthetic scans to balance the dataset resulted in up to a 0.05 increase in the DICE score. Our findings motivate further research on balancing cardiac MRI datasets by using atlas deformation.

External IDs:doi:10.1007/978-3-031-66958-3_26