Fast, Robust and Accurate Cardiac MRI T1 Mapping with Deep Neural Networks
Jeelani, Gulam Mohammed Haras, Electrical Engineering - School of Engineering and Applied Science, University of Virginia
Weller, Daniel, EN-Elec/Computer Engr Dept, University of Virginia
Heart disease is the leading cause of death worldwide, claiming millions of lives each year. Cardiac imaging plays a central role in diagnosis, severity assessment, and treatment guidance of heart disease. Cardiac Magnetic Resonance imaging (CMR) is a popular choice among many other imaging techniques due to its ability both to provide excellent soft-tissue contrast images and to detect different heart diseases. A unique ability of CMR is to quantify small changes in tissue parameters caused by the disease that are difficult to identify visually using other techniques. Although several technological advances have occurred in quantitative CMR, important challenges remain. This thesis addresses some of these challenges of quantitative CMR.
The first contribution of this thesis is to enable fast imaging and fast high-quality reconstruction of CMR data. One of the major limitations of CMR is the slow acquisition speed, due to which high spatial and temporal image resolution is limited by cardiac motion. Reducing the amount of acquired data can accelerate the acquisition; however, the image reconstruction algorithm needs to be able to remove the resulting artifacts. Traditional advanced reconstruction algorithms require hand-crafted priors and time-consuming hyper-parameter tuning. Deep learning-based image reconstruction algorithms provide high-quality images without the need for model-based priors and tuning. In this thesis, a deep recurrent convolutional neural network is applied to reconstruct images from undersampled multicoil data for quantitative CMR parameter mapping. The network does not require tuning and reconstructs higher quality images than the traditional model-based reconstruction algorithm.
The second contribution minimizes the susceptibility of the quantitative parameter estimation process to the noise in the reconstructed images. Conventionally, a voxel-wise regression to a known nonlinear model extracts the tissue parameters. Since the current voxel-wise model does not account for noise and other artifacts, it is not suited for postprocessing images reconstructed from highly undersampled acquisitions. Instead, a data-driven postprocessing approach is proposed to extract the tissue parameters from reconstructed images. This proposed approach provides more accurate tissue parameter values and higher quality tissue parameter maps that are robust to artifacts in the reconstructed images.
In the final contribution of this thesis, the proposed reconstruction and postprocessing methods are applied to detect abnormal tissue parameter values in a hypertrophic cardiomyopathy (HCM) patient. The results of this pilot evaluation suggest the proposed technique does not have a significant bias towards predicting the abnormal tissue parameter values when compared against the conventional technique. This evaluation motivates future investigations validating these data-driven methods in patients with heart diseases, such as HCM.
PHD (Doctor of Philosophy)
Deep Learning, Cardiac MRI, T1-mapping