Unsupervised Domain Adaptation via Variational Autoencoder with Explicit and Implicit Feature Alignment for Cardiac Segmentation

Late gadolinium enhancement (LGE) cardiac magnetic resonance (CMR) is of vital importance in the diagnosis of myocardial infarction (MI). However, the LGE image annotations are limited, it is urgently needed to transfer the learned knowledge in other modalities. We propose an unsupervised domain adaptation (UDA) algorithm based on variational autoencoder (VAE) with a Combination of Explicit and Implicit Alignment of feature distributions (VCEIA). First, we re-derive the Kullback-Leibler (KL) divergence of the features extracted in the source and target domains as a measure of distribution discrepancy. Secondly, for explicit category-level distribution alignment, we use contrast learning to reduce the distance between prototypes of the same category in the source and target domains. Finally, we add a domain discriminator to the output space to implicitly align the feature distributions. Extensive experiments were conducted to validate the effectiveness of VCEIA on the MS-CMRSeg 2019 dataset, which showed competitive results.