X-Ray Scatter Correction by Deep Semi-supervised Learning of Simulated Projections with Beam-hole Array

Abstract

In nondestructive testing and evaluation using cone-beam X-ray computed tomography (CBCT), X-ray scattering is a major cause of degrading the quality of CT images. To correct the negative effects of scattering, recent studies focused on the deep convolutional neural network (DCNN) for fast and accurate correction, and they trained the DCNNs using training data obtained by X-ray simulation. However, such approaches have a drawback where accurately simulating the physical behavior of Xrays requires considerable computational time. Moreover, it is unclear whether the simulated training data are effective to real measured data. To address these problems, we propose a new scatter correction method based on deep semi-supervised learning that can be applied to real measurements. Our training data are incomplete transmission images in which primary X-ray intensities are obtained for a part of pixels. We introduce a data augmentation method based on X-ray radiology, which allows us to significantly reduce training data in our semi-supervised learning. Through experiments using simulated data, this paper shows that our semi-supervised method using a few training data works as effectively as a baseline fully supervised method.