Authors: Dimitrios Bellos (University of Nottingham), Andrew French (University of Nottingham), Mark Basham (Diamond Light Source Ltd),Tony Pridmore (University of Nottingham)

Abstract: Recently there have been a number of approaches which develop a semantic segmentation of X-ray Computed Tomography (CT) scans. In most of these cases, high-quality (HQ) tomography reconstructions are available as input; however, in practice this is not always available. When the X-ray exposure time of the imaged object must be limited to reduce the radiation exposure, or to capture time-critical events, noisy, low-quality (LQ) tomograms will be attained, and as a result these can be more difficult to segment. Here, we address the case of time-resolved volumetric tomography data collection (4D datasets), where multiple LQ tomograms with small number of projections are collected, specifically to investigate the process of industrial corrosion at the Diamond Light Source (DLS) synchrotron. Fortunately however, it is common practice to collect HQ tomograms with a large number of projections before or after the time-critical portion of the experiment. In this paper, we propose an end-to-end network that can learn to denoise and segment the reconstructions of the LQ tomograms, using the representative fully segmented HQ tomogram to train the network. Our single network is able to offer two different desired outputs while only training once, with the denoised output improving the accuracy of the final segmentation. Our method is able to outperform state-of-the-art methods in both tasks of segmentation and denoising. We also make our datasets as well as the code publicly available.

Comments: Presented at BMVC 2019: Workshop on Workshop on Visual Artificial Intelligence and Entrepreneurship (VAIE2019), Cardiff, UK.
Cite as: Paper (PDF): VAIE2019_1.pdf