Unsupervised Physics-Inspired Deep Learning Network with Application to Dental Computed Tomography Image Restoration

Medical image restoration, in particular X-ray computed tomography, remains, to date, an open challenge. A trade-off between the X-ray doses and the image resolution is necessary in medical routine. Thus, many investigations focused on dedicated image processing techniques to allow high-resolution imaging at low doses. While the initial methods were model-based, currently, deep learning architecture-based approaches provide state-of-the-art results in this domain. Recently, physics-aware approaches combining quantum mechanics principles and convolutional neural networks have been proposed to tackle challenging inverse problems, with promising results. These approaches rely on supervised architectures and have been only tested on photographic images. This paper proposes an unsupervised version of a quantum mechanics-based convolutional neural network with a special focus on dental computed tomography image restoration. Comparisons with the initial supervised network and state-of-the-art image restoration methods confirm the promising performance of the proposed unsupervised technique.