Improving Spectral CT Image Quality Based on Channel Correlation and Self-Supervised Learning

Photon counting spectral computed tomography (PCCT) can produce reconstructed attenuation maps in different energy channels, reflecting the energy properties of the scanned object. Due to the limited photon numbers of each energy channel and the nonideal detector response, the reconstructed images usually contain considerable noise. With the development of the deep learning (DL) technique, different DL-based models have been proposed for noise reduction in CT. However, most of the models require paired datasets for training, which are rarely available in practical imaging procedures. Inspired by the structural similarities of each channel's reconstructed image, we proposed a self-supervised learning based PCCT image enhancement framework via multi-spectral channels (S$^{2}$MS). In the S$^{2}$MS framework, both the input and output labels are noisy images. Specifically, one single energy channel image was used as output. The other channel images and a full-energy image were used as input to train the network, which can fully use the spectral data information without extra cost. Experiments on simulated and real noisy data demonstrate that the proposed S$^{2}$MS model can suppress noise and preserve details more effectively than traditional DL models, and has the potential to improve PCCT image quality in clinical applications.