Application of machine-learning models to improve the image quality of photon-counting CT images

Abstract

Abstract X-ray computed tomography (CT) has been widely used in medical diagnostic imaging. However, conventional, energy-integrated CT requires a high radiation dose and can only provide monochromatic images that cannot eliminate various artifacts. In contrast, photon-counting CT (PC-CT) provides low-dose multicolor CT imaging, which enables the identification of multiple contrast agents. However, in the PC-CT system, the lack of photon statistics, which is also caused by image reconstruction in the limited energy band, severely affects the image quality. In this study, we applied three types of machine-learning (ML) techniques to improved the image quality of PC-CT, that is, dictionary learning, U-Net, and Noise2Noise. These ML models were trained using low- and high-dose image pairs created in simple steps. The trained ML models were applied to simulated data, and experimental PC-CT images of contrast agents used in clinical practice. Consequently, in the simulated data, the peak signal-to-noise ratio (PSNR) value improved from 21.3 for the input to 26.6, 33.3, and 30.1 for dictionary learning, U-Net, and Noise2Noise, respectively. Furthermore, in the actual PC-CT images, we successfully reproduced PC-CT images with high PSNR, which enabled simultaneous imaging of multiple contrast agents with improved accuracy of concentration estimation. As a future perspective, we will develop a processing technique that can be applied to in vivo CT images.