Abstract
This thesis presents a series of interrelated studies I conducted to contribute to the MARS Spectral Photon Counting CT (SPCCT) research, development, and commercialisation. There are two main parts: the first focuses on material decomposition, and the second focuses on applying deep learning image-denoising techniques.
Firstly, I co-developed a new MARS material decomposition software known as MARS FastMD. As the name implies, the original motivation was to create a faster material decomposition software to support the current demand of the MARS team. During the past few years, this software has been the primary material decomposition tool for the MARS team. The software is an all-in-one package that includes all necessary steps to perform and evaluate material decomposition, including basis vector calculations, the material decomposition itself, and evaluation tools. In addition, it uses a direct matrix inversion algorithm to perform the material decomposition with vectorised operations, thus significantly decreasing the software run time.
I also present a new method to generate probability maps for material images. The probability maps provide end-users extra information on the per-voxel probability values of a material identified in a voxel. This additional information gives users greater confidence in their MD results. Another valuable use case for probability maps is as an evaluation metric to compare different MD algorithms.
In the second part, I explored supervised and self-supervised neural network models to perform image denoising on MARS images. Due to the limited availability of training datasets, I employed a synthetic-to-real transfer learning method for the supervised model. In addition, I introduced a new SVD method to generate custom virtual phantoms to train a Fully-Convolutional DenseNet for image denoising.
The self-supervised model utilises the Noise2Inverse framework with Mixed-Scale DenseNet. To compare the effect of this framework on different reconstruction algorithms, I also developed a new reconstruction software using a deep learning framework enabling FDK and SIRT reconstructions. Furthermore, I applied the self-supervised framework to the SIRT reconstruction and the MARS commercial reconstruction software. Finally, I introduced a custom fusion network and tested its denoising capability and structural preservation.
In conclusion, the outputs of this thesis include two software libraries, a new method for material decomposition probability mapping, and a new method to generate custom virtual phantoms. In addition, this thesis has also demonstrated the application of two different neural network models on MARS SPCCT image denoising.