Geometric Artifact Evaluation of X-ray Computed Tomography Images Based on Convolutional Neural Network

Abstract

Computed Tomography (CT) has been used in many fields. Practical misalignment of the actual CT system causes geometric artifacts in the reconstructed images, which severely degrades image quality. Geometric artifact evaluation provides a reliable basis for subsequent geometric artifact correction. Some characteristics of images, such as entropy and sharpness, are often used to measure the severity of geometric artifacts, but they are limited in generality and accuracy. Convolutional neural network (CNN) has excellent image feature learning capabilities and is well used in image processing. This paper explores the network structure suitable for the evaluation of geometric artifacts in CT images. We select three commonly used networks LeNet-5, VGG16 and AlexNet. The datasets of three kinds of phantoms are constructed using simulation and actual scanning results. The three networks are trained and tested separately on the three kinds of datasets. Experimental results show that all three CNN models can evaluate the existence of geometric artifacts in CT images. The AlexNet network achieves the best classification evaluation performance with an average classification accuracy of 0.961, with the smallest average loss and the shortest training time.