Image improvement in digital tomosynthesis (DTS) using a deep convolutional neural network

Abstract

Digital tomosynthesis (DTS) is a geometric tomography technique using a limited-angle scan. It has been popularly used in both medical and industrial x-ray imaging applications. DTS provides the tomographic benefits of computed tomography with reduced dose and time. However, conventional DTS reconstruction based on the computationally cheap filtered back-projection (FBP) method typically produces poor image quality due to limited angular samplings. To overcome these difficulties, iterative reconstruction methods are often used in DTS reconstruction as they have the potential to provide multiplanar images of higher quality than conventional FBP-based methods. However, they require enormous computational cost in the iterative process, which remains an obstacle to practical applications. In this study, we propose a method for effectively reducing limited-angle artifacts in conventional FBP reconstruction, using a state-of-the-art deep learning scheme with a convolutional neural network. Our results indicate that the proposed DTS reconstruction method effectively minimized limited-angle artifacts, thus improving image performance in DTS, and that further it provided good image quality in both sagittal and coronal views (as in computed tomography) as well as in axial view.