Unsupervised domain adaptation for cardiac MRI segmentation via adversarial learning in latent space

Late gadolinium enhancement (LGE) cardiac magnetic resonance (CMR) imaging is crucial for visualizing myocardial infarction (MI), with accurate segmentation of the ventricles and myocardium being essential for effective MI treatment. However, due to the complex myocardial structure and the limited availability of pixel-level annotations in LGE CMR images, accurate segmentation using supervised deep learning methods remains challenging. To address this, we propose an unsupervised domain adaptation framework for LGE CMR segmentation, utilizing CMR images from other modalities. First, we transform balanced Steady-State Free Precession (bSSFP) CMR images, which have abundant annotations, into LGE-like images using an enhanced CycleGAN. This CycleGAN incorporates an adversarial sample mining technique in the latent space to improve the quality of synthetic images. Next, we modify the nnU-Net architecture by introducing non-local blocks to train on these synthetic images, enabling precise segmentation of the myocardium and ventricular regions. We evaluate our method on the MS-CMRSeg 2019 dataset and MyoPS 2020 dataset, achieving an average Dice score of 88.0 % and 82.6 % respectively. Our experimental results demonstrate superior performance compared to state-of-the-art methods. The code for our approach is available at https://github.com/Lucarqi/Adv-CycleGAN.