A deep-learning-based workflow for reconstructing and segmenting challenging sets of time-resolved X-ray micro-computed tomography data

Abstract

We present a deep-learning-based software pipeline for reconstructing and segmenting large sets of time-resolved micro-computed tomography (µCT) image data. We construct and train a convolutional neural network (CNN) to consistently, rapidly, and autonomously segment the time-resolved tomography data. The preceding CT reconstruction steps are parametrized for optimal image quality for segmentation. We demonstrate how to discriminate materials with similar radiographic densities in the presence of different media, such as air and water. Our approach can be used out of the box for similar µCT data or adapted to any similarly challenging 3D image data by retraining the neural network.

Keywords

• Time-resolved computed tomography
• Image segmentation
• CT reconstruction
• Convolutional neural networks
• Phase retrieval