Evaluation of deep-learning image reconstruction for photon-counting spectral CT : A comparison between image domain- and projection domain-denoising

A promising new technology in medical imaging is photon-counting detectors (PCD). Itcould allow for images with higher resolution, less noise, improved material decomposi-tion while possibly reducing radiation exposure for patients. Recently, the possibility touse deep-learning denoising in tandem with PCD to increase image quality is starting tobe investigated. In this report we use a variety of standard image quality metrics suchas MSE, SSIM and MTF, on different image phantoms, to evaluate two ways of imple-menting neural networks in the reconstruction process: in the image domain and in thesinogram domain. We show that implementing the network in the image domain seemsto be the most promising choice to increase image quality, observing higher contrast,reduced noise and smaller errors than for the sinogram domain network. We also discusswhy this might be the case. Additionally, we study the effects of optimizing the networksand how well the neural networks generalize to types of phantoms other than the onesthey were trained on