SMOD - Data Augmentation Based on Statistical Models of Deformation to Enhance Segmentation in 2D Cine Cardiac MRI

Deep learning has revolutionized medical image analysis in recent years. Nevertheless, technical, ethical and financial constraints along with confidentiality issues still limit data availability, and therefore the performance of these approaches. To overcome such limitations, data augmentation has proven crucial. Here we propose SMOD, a novel augmentation methodology based on Statistical Models of Deformations, to segment 2D cine scans in cardiac MRI. In brief, the shape variability of the training set space is modelled so new images with the appearance of the original ones but unseen shapes within the space of plausible realistic shapes are generated. SMOD is compared to standard augmentation providing quantitative improvement, especially when the training data available is very limited or the structures to segment are complex and highly variable. We finally propose a state-of-art, deep learning 2D cardiac MRI segmenter for normal and hypertrophic cardiomyopathy hearts with an epicardium and endocardium mean Dice score of 0.968 in short and long axis.