Supervisors: Dr. Nick Polydorides (University of Edinburgh), Dr. Dimitris Kamilis (University of Edinburgh) and Prof. Ender Konukoglu
Coronary heart disease is the leading cause of death in developed countries and coronary artery calcification is an important predictor of cardiac events. We present two deep learning approaches that could help improve diagnosis of coronary artery calcification with limited angle spectral CT. To this end, we embed two unrolled iterative primal-dual optimisation schemes into neural networks and produce material separated reconstructed images. Both methods are compared and evaluated on simulated random ellipse and human thoracic phantoms containing atherosclerotic plaque with 3 and 5 different materials. We further investigate the influence of different neural network hyperparameters and compare against a classical material decomposition and reconstruction algorithm. The method works very well for separating and reconstructing 3 materials, with structural similarities of over 99\%. We also show very good performance in separating 5 materials for the medical phantoms with structural similarities of 98\% and higher, outperforming the classical approach for the chosen settings. Apart from material decomposition and good quality reconstruction, our method enables a highly reduced reconstruction time compared to other iterative reconstruction algorithms and renders cumbersome regularisation redundant.