RecDNN: deep neural network for image reconstruction from limited view projection data

Abstract

Improving the quality of reconstruction, even in a limited projection view, has become one of the prime objectives in computed tomography. Projection data over 180 $$^{\circ }$$ are always not practicable in many practical applications including medical imaging. Iterative approaches, such as algebraic reconstruction techniques (ART), are useful in the limited view scenarios; however, these approaches resulted in the streaking artifacts. Deep learning has paved a path to achieve high efficiency and accuracy for reconstruction with limited view projection data. Convolutional neural networks such as U-Net, residual neural networks and adversarial neural networks are being widely used for image reconstruction. U-Net and residual neural network for limited projection views have been found compute intensive and also produces a noisy image. The convergence of the adversarial neural network in case of the inverse problem is very slow; hence, the training is compute intensive. A deep neural network for image reconstruction (RecDNN) is proposed in this manuscript for a limited view case. The proposed architecture is designed to give reconstruction with minimal artifacts. The objective is to give a good quality of textural and contrast information about the reconstructed object. The image reconstruction has been addressed as inverse problem as well as optimization problem. The stochastic gradient method is applied to achieve image reconstruction. The proposed approach has been compared to the traditional transform-based approaches as well as the current deep learning methods for image reconstruction. Experimental results show that the proposed architecture is performing better than the existing approaches, as well as the state-of-the-art architectures available for image reconstruction. Experimental results also show that for the same number of projection views, the proposed architecture reconstructs the object with higher quality as compared to existing architectures for image reconstruction.