Segmentation of Cardiac Magnetic Resonance Imaging (MRI): Generative Adversarial Networks and Asynchronous Federated Learning
Abstract: Cardiac magnetic resonance (CMR) imaging allows non-invasive, non-ionizing assessment of cardiac function and anatomy and is especially beneficial in patients with congenital heart disease (CHD). Current clinical workflow mostly relies on manual analysis of CMR images. Segmentation of different chambers, mainly the left and right ventricles, from CMR datasets is an essential step for calculation of clinical indices such as ventricular volume and ejection fraction. We present deep learning methods and deformable models that can automatically segment CMR images and improve the accuracy and robustness. Training these models requires large annotated datasets, which are not readily available particularly for CHD patients. To mitigate this issue, we devised a novel method that uses a generative adversarial network (GAN) to synthetically augment the training dataset via generating synthetic CMR images and their corresponding chamber segmentations. Our methods outperform the existing techniques presented in different MICCAI challenges and commercially available segmentation software. Finally, we argue that distributed learning methods that train locally but build global models can increase the access by preserving the privacy. We discuss asynchronous federated learning methods that provide high precision while significantly reducing the data transmitted during client-server interactions.