Scatter correction of cone-beam CT using a deep residual convolution neural network (DRCNN)

Abstract

Scatter correction is an essential technique to improve the image quality of cone-beam CT (CBCT). Although different scatter correction methods have been proposed in the literature, a standard solution is still being studied due to the limitations including accuracy, computation efficiency and generalization. In this paper, we propose a novel scatter correction scheme for CBCT using a deep residual convolution neural network (DRCNN) to overcome the limitations. The proposed method combines the deep convolution neural network (CNN) and the residual learning framework (RLF) to train the mapping function from the uncorrected image to the corrected image. Two residual network modules (RNMs) are built based on the RLF to improve the accuracy of the mapping function by strengthening the propagation of the gradient. The dropout operations are applied as the regularizer of the network to avoid the overfitting problem. The RMSE of the corrected images reconstructed using the DRCNN is reduced from over 200 HU to be about 20 HU. The structural similarity (SSIM) is slightly increased from 0.95 to 0.99, indicating that the proposed scheme maintains the anatomical structure. The proposed DRCNN has a higher accuracy of scatter correction than the networks without the RLF or the dropout operations. The proposed network is effective, efficient and robust as a solution to the CBCT scatter correction.