Enhancement of Diagnostic Information in T1-weighted Contrast Enhanced MRI using Deep Back-Projection based Super-Resolution

Deep learning based magnetic resonance (MR) image super-resolution (SR) using a convolutional neural network (CNN) is gaining popularity very rapidly. Since MR images generally suffer from low resolution, SR reconstruction is very useful for clinical analysis. Most of the existing SR network architectures are based on absolute error minimization. However, in medical images, preservation or enhancement of diagnostic quality of SR reconstructed images is an essential requirement. In this paper, we propose a deep back-projection network architecture to improve diagnostic information in SR reconstruction. We have also introduced a novel diagnostic information distortion measurement parameter based on the normalized spectral energy difference. For performance evaluation, we consider a T1-weighted contrast enhanced MRI pathological dataset. From the experimental results, we observed that the proposed technique outperforms well-known existing in terms of preservation of fine details, edges, and diagnostic information, compared to some of the recent SR methods.