Generative Adversarial Networks for X-Ray Computed Tomography
posterposted on 30.11.2020, 17:09 by Emilien Valat
X-Ray computed tomography is a widely used, non-destructive imaging technique that produces cross-sectional images of bodies sensitive to X-Ray. Inter alia, it relies on exhaustive sampling of the attenuation properties of the scanned material and advanced reconstruction processes. However, acquisition can be toxic for humans or limiting for exotic geometries, as intense X-Ray exposure can lead to cancers during in-vivo diagnosis and experiments chambers have a fixed size that might limit the information gathering process for certain objects. Since sparse data from incomplete scans is yet to be compensated by adequate aftertreatment, we have decided to use deep-learning techniques to extract information on additional modalities to generate missing data in the acquisition.
In many routine diagnoses, prior knowledge about the scanned object is often known. Whether it is computer-assisted design drawings or anatomical models, the availability of information regarding the shape of the test sample has led us to look for an acquisition process that minimises object sampling and maximises data harnessing on a known modality. After an introductory period of looking for the suitable architecture and publishing negative results, our exploration of deep generative models has led us to a unique design, one that combines unsupervised feature extraction with graphical models, use of these features for image generation with likelihood-free networks and a constrained optimisation problem to generate high-resolution acquisitions. This model translates our optimal understanding of the problem and an initial analysis suggests the feasibility of our process. Should the concept be promising, many challenges are yet to be addressed: accurate database constitution, efficient training items generation, thorough hyperparameters optimisation and delicate experimentations. As such, these are the next milestones in this investigation. Over the course of the next year, we are determined to deliver a method that is not only novel, but useful to many research fields.