Machine Learning-Driven Dynamic Four-Dimensional X-Ray Computed Tomography Reconstruction

Hyojin Kim | 20-FS-010

Project Overview

Dynamic computed tomography (DCT) refers to image reconstruction of moving or non-rigid objects over time while x-ray projections are acquired over a range of angles. We present two novel reconstruction approaches to reconstruct dynamic scenes as time-varying sequence of 3D volumes (D4DCT). The supervised method leverages convolutional neural networks by training input sinogram data and its ground truth image sequence, to reconstruct an initial scene with sequence of motion warping fields. The second reconstruction method leverages implicit neural representations with a parametric motion fields to reconstruct dynamic scenes as time-varying sequence of 3D volumes without any need of training data. The proposed method is an end-to-end, coarse-to-fine, optimization approach without the need for any training data by minimizing the measured x-ray sinogram and the forward projected sinogram of the estimated scene. Furthermore, we provide challenging D4DCT datasets where the sinogram data represent a time-varying object deformation to demonstrate damage evolution under mechanical stresses. The provided datasets enable training and quantitative evaluation of the data driven machine learning approaches for the D4DCT reconstruction.

Mission Impact

The outcome of this project will have a significant impact on multiple missions and areas of programmatic interest in geoscience, additive manufacturing (AM) in materials science, and other non-destructive evaluation CT applications. One application area is in-situ CT imaging to study material deformation and other characterization of geometrically complex AM parts under mechanical loading, and compaction of confined granular materials. In that study, the material scientists need to capture in-situ porosity and microstructural damage evolution of materials (e.g., crack and fatigue analysis, failure and void region detection) subjected to continuous mechanical and thermal loads using non-interrupted CT. The proposed dynamic 4D CT capabilities are crucial to enable such non-interrupted CT imaging setups as well as to advance AM parts. The proposed approach will be beneficial to a number of data science applications regarding temporal data analysis and highly ill-posed inverse problems.

Publications, Presentations, and Patents

Dynamic 4DCT Reconstruction using Neural Representation-based Optimization, ROI submitted, IL Number: IL-13625.

LLNL D4DCT Datasets: Dynamic 4DCT Datasets using MPM-based Deformation, to appear at DSI's Open Data Initiative, LLNL-MI-81669.