Comparison of Semi- and Un-Supervised Domain Adaptation Methods for Whole-Heart Segmentation

Quantification of heart geometry is important in the clinical diagnosis of cardiovascular diseases. Changes in geometry are indicative of remodelling processes as the heart tissue adapts to disease. Coronary Computed Tomography Angiography (CCTA) is considered a first line tool for patients at low or intermediate risk of coronary artery disease, while Coronary Magnetic Resonance Angiography (CMRA) is a promising alternative due to the absence of radiation-induced risks and high performance in the evaluation of cardiac geometry. Yet, the accuracy of an image-based diagnosis is susceptible to the quality of volume segmentations. Deep Learning (DL) techniques are gradually being adopted to perform such segmentations and substitute the tedious and manual work performed by physicians. However, practical applications of DL techniques on a large scale are still limited due to their poor adaptability across modalities and patients. Hence, the aim of this work was to develop a pipeline to perform automatic heart segmentation of multiple cardiac imaging scans, addressing the domain shift between MRs (target) and CTs (source). We trained two Domain Adaptation (DA) methods, using Generative Adversarial Networks (GANs) and Variational AutoEncoders (VAEs), following different training routines, which we refer to as un- and semi- supervised approaches. We also trained a baseline supervised model following state-of-the-art choice of parameters and augmentation. The results showed that DA methods can be significantly boosted by the addition of a few supervised cases, increasing Dice and Hausdorff distance metrics across the main cardiac structures.