Self-supervised Projection Denoising for Low-Dose Cone-Beam CT

We consider the problem of denoising low-dose xray projections for cone-beam CT, where x-ray measurements are typically modeled as signal corrupted by Poisson noise. Since each projection view is a 2D image, we regard the lowdose projection views as examples to train a convolutional neural network. For self-supervised training without ground truth, we partially blind noisy projections and train the denoising model to recover the blind spots of projection views. From the projection views denoised by the learned model, we can reconstruct a high-quality 3D volume with a reconstruction algorithm such as the standard filtered backprojection. Through a series of phantom experiments, our self-supervised denoising approach simultaneously reduces noise level and restores structural information in cone-beam CT images.

I. INTRODUCTION

Onboard cone-beam computed tomography (CBCT) is commonly used for patient setup and adaptive replanning in radiation therapy. Although CBCT is capable of quickly producing on-treatment patient anatomy, the repeated utilization of CBCT has increased concern about the risk of the radiation dose [1]. To reduce radiation exposure of patients, low-dose protocols with tube current modulation and lower tube voltage have been applied to CBCT imaging [2], [3]. However, low-dose CBCT protocols also decrease signal-to-noise ratio (SNR) in x-ray measurements, which may adversely affect their clinical usefulness [4], [5].