Deep Learning Based Filter with Back-Projection Operator for CT Reconstruction

Abstract

This study analyzes the inherent limitations of traditional filtering methods in the filtered back projection (FBP) algorithm for CT image reconstruction. The main goal is to explore the feasibility of using convolutional neural networks (CNN) to replace traditional filters, thereby improving the quality of CT image reconstruction. This study introduces a novel deep learning back projection (DLBP) framework that combines a CNN with a back projection operator. This framework takes the advantage of CNN as deep learning filter in image processing. And keep the conventional back projection operator for transformation from sinogram domain to image domain. Unlike the other study, the back projection operator also involved the training process. While the CNN filter and the fixed back projection operator together mapping the sinogram and target image, the CNN part automatically becoming a data driven neural network filter. The materials used in this study include a public dataset from Shanghai Medical College of Fudan University, which contains 27,800 high-resolution chest and abdominal CT slices from 30 patients. These image data are preprocessed and the corresponding sinograms are generated, then sinograms used as inputs, while the original images are used as targets to form dataset pairs. In the study, the DLBP method was compared with the conventional FBP algorithm using image evaluation metrics such as mean square error (MSE), structural similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR). The results show that the DLBP method performs significantly better than the FBP algorithm in noisy environments, achieving lower MSE, higher SSIM, and better PSNR values. The results highlight the potential of the DLBP framework in enhancing the quality of CT image reconstruction and suggest future research directions, namely validating the framework on larger datasets, exploring complex geometries, and leveraging advanced hardware to improve performance.