DUG-RECON: A Framework for Direct Image Reconstruction Using Convolutional Generative Networks

Abstract

This article explores convolutional generative networks as an alternative to iterative reconstruction algorithms in medical image reconstruction. The task of medical image reconstruction involves mapping of projection domain data collected from the detector to the image domain. This mapping is done typically through iterative reconstruction algorithms which are time consuming and computationally expensive. Trained deep learning networks provide faster outputs as proven in various tasks across computer vision. In this work, we propose a direct reconstruction framework exclusively with deep learning architectures. The proposed framework consists of three segments, namely, denoising, reconstruction, and super resolution (SR). The denoising and the SR segments act as processing steps. The reconstruction segment consists of a novel double U-Net generator (DUG) which learns the sinogram-to-image transformation. This entire network was trained on positron emission tomography (PET) and computed tomography (CT) images. The reconstruction framework approximates 2-D mapping from the projection domain to the image domain. The architecture proposed in this proof-of-concept work is a novel approach to direct image reconstruction; further improvement is required to implement it in a clinical setting.