Principal Investigator: Dr Xue Feng
University of Virginia, USATags: 52894, cardiac MRI, deep learning, featured, incremental learning, myocardial perfusion, scar quantification, transfer learning
The aim of this study is to use the cardiac MRI data from UK Biobank, including the cine, native T1 and tagging images and the correspondingly defined regions-of-interest (ROIs) to assist the segmentation of other types of cardiac MRI, such as myocardial first-pass perfusion and late gadolinium enhancement (LGE) images. The biggest advantage of the UK Biobank is the large data size and diversity. Therefore, convolutional neural networks (CNNs) can be trained to take this advantage and perform segmentation tasks with very high accuracy and reliability, as proven in other studies. Perfusion and LGE images can provide additional clinical information in diagnosis of cardiac diseases and are often used in clinical environments. Further quantification from these images to extract clinically relavant biomarkers, as compared to simple human rating, has the potential to improve diagnostic accuracy and robustness and reduce human bias. We’ve shown that segmentation of these images can improve the quantification process by accurately defining and using the ROIs. However, there are no available sources for a large scale perfusion and LGE images to train the segmentation networks. In this project, we aim to leverage the UK Biobank data to assist the segmentation of perfusion and LGE images with only a small local dataset and use the trained network to improve the quantification process and ultimately the accuracy and reliability of diagnosis.
Last updated Nov 11, 2019