Learning-Based Super-Resolution Applied to Dental Computed Tomography. (2019) IEEE Transactions on Radiation and Plasma Medical Sciences, 3 (2). 120-128.

Available super-resolution techniques for 3D images are either computationally inefficient prior-knowledge-based iterative techniques or deep learning methods which require a large database of known low-and high-resolution image pairs. A recently introduced tensor-factorization-based approach offers a fast solution without the use of known image pairs or strict prior assumptions. In this article this factorization framework is investigated for single image resolution enhancement with an off-line estimate of the system point spread function. The technique is applied to 3D cone beam computed tomography for dental image resolution enhancement. To demonstrate the efficiency of our method, it is compared to a recent state-ofthe-art iterative technique using low-rank and total variation regularizations. In contrast to this comparative technique, the proposed reconstruction technique gives a 2-order-of-magnitude improvement in running time -2 minutes compared to 2 hours for a dental volume of 282×266×392 voxels. Furthermore, it also offers slightly improved quantitative results (peak signalto-noise ratio, segmentation quality). Another advantage of the presented technique is the low number of hyperparameters. As demonstrated in this paper, the framework is not sensitive to small changes of its parameters, proposing an ease of use.

A volumetric non-blind single image super-resolution technique using tensor factorization has been recently introduced by our group. That method allowed a 2-order-of-magnitude faster high-resolution image reconstruction with equivalent image quality compared to state-of-the-art algorithms. In this work a joint alternating recovery of the high-resolution image and of the unknown point spread function parameters is proposed. The method is evaluated on dental computed tomography images. The algorithm was compared to an existing 3D super-resolution method using low-rank and total variation regularization, combined with the same alternating PSF-optimization. The two algorithms have shown similar improvement in PSNR, but our method converged roughly 40 times faster, under 6 minutes both in simulation and on experimental dental computed tomography data.

No abstract