Robust Cardiac MRI Segmentation with Data-Centric Models to Improve Performance via Intensive Pre-training and Augmentation

Shizhan Gong, Weitao Lu, Jize Xie, Xiaofan Zhang, Shaoting Zhang, Qi Dou

Abstract: Segmentation of anatomical structures from Cardiac Magnetic Resonance (CMR) is central to the non-invasive quantitative assessment of cardiac function and structure, and deep-learning-based automatic segmentation models prove to have satisfying performance. However, patients’ respiratory motion during the scanning process can greatly degenerate the quality of CMR images, resulting in a serious performance drop for deep learning algorithms. Building a robust cardiac MRI segmentation model is one of the keys to facilitating the use of deep learning in practical clinic scenarios. To this end, we experiment with several network architectures and compare their segmentation accuracy and robustness to respiratory motion. We further pre-train our network on large publicly available CMR datasets and augment our training set with adversarial augmentation, both methods bring significant improvement. We evaluate our methods on the cine MRI dataset of the CMRxMotion challenge and obtain promising performance for the segmentation of the left ventricle, left ventricular myocardium, and right ventricle.