Deep neural network for beam hardening artifacts removal in image reconstruction

Image reconstruction with limited angles projection data is a challenging task in computed tomography (CT). The amount of radiation associated with CT induces health implications to the patient. Besides, image reconstruction with limited-angles projection data distorts the image, thus emasculating the efficiency of diagnosis. Also, the poly-chromatic nature of the X-ray adds beam-hardening artifacts in the reconstruction. The state-of-the-art approaches available in the literature have proposed the solutions for beam-hardening artifacts correction in full span computed tomography. Most of the solutions are hardware based and need extra hardware to remove the beam hardening artifacts. The present manuscript proposes artificial intelligence based software solution for the beam hardening artifacts removal. This manuscript has presented a cascaded encoder-decoder architecture named cascaded deep neural network for image reconstruction (CDNN). The CDNN architecture has convolution neural network blocks that include convolution layers, rectified linear units ReLU, and batch normalization layers. The network has skip-connections for better learning of features between input and output. The network has been designed as a forward model. The stochastic gradient descent optimization method has been used for training the network. Image reconstructed from Fourier transform-based approach has been used as a prior. A novel approach for reduction of beam-hardening artifacts in case of limited-angles computed tomography using CDNN has been presented. The proposed approach is comparable to other hardware/software solutions for aforesaid purpose and does not require any extra hardware. The proposed approach has improved the image quality as compared to U-Net and the other state-of-the-art methods. It has been found from the experiments that the CDNN suppresses the artifacts and improves the reconstruction. The performance of the proposed CDNN has been tested with real-life data having beam hardening artifacts. It has been observed that the CDNN has improved the reconstruction quality by reducing streak, ring artifacts, and beam hardening artifacts and also preserving the profound structures.