A deep learning reconstruction framework for X-ray computed tomography with incomplete data

Abstract

As a powerful imaging tool, X-ray computed tomography (CT) allows us to investigate the inner structures of specimens in a quantitative and nondestructive way. Limited by the implementation conditions, CT with incomplete projections happens quite often. Conventional reconstruction algorithms are not easy to deal with incomplete data. They are usually involved with complicated parameter selection operations, also sensitive to noise and time-consuming. In this paper, we reported a deep learning reconstruction framework for incomplete data CT. It is the tight coupling of the deep learning U-net and CT reconstruction algorithm in the domain of the projection sinograms. The U-net estimated results are not the artifacts caused by the incomplete data, but the complete projection sinograms. After training, this framework is determined and can reconstruct the final high quality CT image from a given incomplete projection sinogram. Taking the sparse-view and limited-angle CT as examples, this framework has been validated and demonstrated with synthetic and experimental data sets. Embedded with CT reconstruction, this framework naturally encapsulates the physical imaging model of CT systems and is easy to be extended to deal with other challenges. This work is helpful to push the application of the state-of-the-art deep learning techniques in the field of CT.

A review on medical imaging synthesis using deep learning and its clinical applications

“…, most of the methods in the reviewed literature implement direct image translation from low-dose to full-dose CT while others restore the sinogram using deep learning first, and then reconstruct images from the restored sinogram by FBP. As shown byDong et al, their proposed projection-based method outperformed an image-based method by better reducing downsampling artifacts with higher resolution on object edges 119. A potential reason for this difference is that for image-based methods prediction error is directly observed on the image while for projection-based methods, the error predicted on the sinogram will be compensated for in the reconstruction process, where the final product is a weighted sum of all sinograms.…”

Hydrodynamic and Radiographic Toolbox (HART)

“…reconstructions. More recently, a new deep learning reconstruction framework for CT with incomplete projections utilized tight coupling of the deep learning U-net and FBP algorithm in the domain of the projection sinograms [29]. Indeed, DL methods described have shown tremendous potential in reducing the number of views and retaining excellent reconstruction ability [39,48,85].…”

Artificial intelligence with deep learning in nuclear medicine and radiology

“…The aforementioned networks can also be adapted for other types of inputs and outputs. Some studies on limited angle tomography, for example, choose to perform image restoration in sinogram space ( ) prior to image reconstruction [ 77 , 78 ], although both options are compared for partial-ring PET in [ 79 ], showing better results using image space data. Alternatively, dual imaging modalities such as PET/MRI may use the MRI scan as an additional input to provide anatomical information, helping with the denoising of the PET scan [ 80 ].…”