A 3D Convolutional Neural Network for Denoising of Proton CT

Abstract

Filtered back projection (FBP) is the traditional image reconstruction method in x-ray computed tomography (CT). To further improve the quality of FBP-based image reconstruction, often iterative methods are used. But these methods are slow and computationally expensive. In applications where fast and accurate images are needed, FBP may not be sufficient. We propose a very lightweight 3D convolutional neural network (3D CNN) to improve the image quality of the FBP method with minimal computational cost. 3D CNN can be used in time-sensitive situations where iterative methods are not ideal due to time constraints. The method takes advantage of the 3D-nature of the CT data. Popular denoising CNN-based algorithms have been applied to 2D images and were originally designed to correct the effects of 2D artifacts like jpg compression. These approaches fail to utilize the 3D geometry of the CT data and try to improve the 2D CT images individually. The focus of this work was on training and applying a time-saving method to proton CT head data. Proton CT (pCT) requires low-noise reconstruction of relative stopping power (RSP) for proton beam therapy planning and pretreatment verification, where time is of the essence. The proposed fast method successfully improved peak SNR (PSNR) of FBP pCT images and showed a numerical advantage of PSNR for 3D over 2D approaches. In future work, the implementation of existing popular 2D denoising methods in 3D CNN CT reconstruction will be investigated.