Low‐dose spectral CT reconstruction based on structural prior network

BACKGROUND: The insufficiency of photon counting in spectral CT leads to statistical noise in the reconstructed images and the deviation of material decomposition. Currently, deep learning is an effective method for removing image noise. PURPOSE: However, in actual experiments, the normal-dose spectral CT images used as network labels are still accompanied by some noise and artifacts, which is unfavorable in the network training and cause the output images to have residual noise. METHODS: To solve this problem, we propose a structural prior network (SP-Net) for low-dose spectral CT image denoising, which can still achieve denoise well when using noisy labels. Inspired by the compressed sensing framework, the structure information of the prior images is integrated into the loss strategy of the network model to guide the network training. Therefore, the network training depends not only on the supervision of labels but also on the prior structure information, to reduce the adverse effects on the noisy labels and improve image quality. RESULTS: We demonstrate the effectiveness of the proposed method in the simulations and experiments. The results show that the proposed method can eliminate the adverse effects of noise labels, remove the noise, and preserve the image structure well. CONCLUSIONS: This method can solve the problem of network training using noisy labels of spectral CT, provide references for future researches, and have high application value in medical imaging.