Three-dimensional visualization of X-ray micro-CT with large-scale datasets: Efficiency and accuracy for real-time interaction

Yipeng Yin, Rao Yao, Qingying Li, Dazhong Wang, Hong Zhou, Zhijun Fang, Jianing Chen, Longjie Qian, Mingyue WuComput. Sci. Rev 2026 · DOI

TL;DR

This paper reviews advancements in 3D visualization of X-ray Micro-CT data, focusing on the critical trade-off between accuracy and efficiency for large datasets. It surveys CT reconstruction and volume rendering techniques, highlighting methods that balance these factors and exploring future directions for real-time interaction and digital twin applications.

What Problem Does This Paper Address?

X-ray Micro-CT is increasingly used for detailed material microstructure analysis, but this generates massive datasets. The challenge lies in accurately visualizing these large datasets efficiently, enabling real-time interaction and analysis. Traditional methods struggle with this trade-off, leading to either slow processing or compromised accuracy. This paper addresses this by reviewing and analyzing current and emerging techniques in CT reconstruction and volume rendering that aim to overcome these limitations.

Methods

The paper provides a comprehensive review of techniques in two main areas:

  • CT Reconstruction: It discusses both analytical methods (like Filtered Back Projection - FBP, and Feldkamp-Davis-Kress - FDK) and iterative reconstruction (IR) algorithms. IR methods are highlighted for their ability to reduce noise and artifacts, with a focus on advancements like hybrid-IR, model-based IR (MBIR), and the emerging field of Deep Learning Reconstruction (DLR). DLR is presented as a promising approach for improving image quality, reducing artifacts, and accelerating reconstruction.
  • Volume Rendering: The paper surveys various volume rendering techniques, including Ray Casting, Splatting, 3D Texture-Mapping, and Shear-warp. It emphasizes techniques for enhancing efficiency and quality, such as parallel rendering, GPU acceleration, adaptive sampling, multi-resolution rendering, and advanced lighting models like Cinematic Rendering (CR). CR, which uses physics-based lighting models, is discussed for its ability to produce more realistic visualizations, though it requires higher computational power.

Datasets

Not explicitly discussed in the paper. The paper focuses on reviewing existing techniques and their principles rather than presenting new experimental results based on specific datasets.

Key Insights and Ideas

  • The Accuracy-Efficiency Dilemma: The core challenge in Micro-CT visualization is balancing the need for high accuracy in representing fine details with the demand for efficient processing of massive datasets for real-time interaction.
  • Evolution of Reconstruction: CT reconstruction has moved from analytical methods to more sophisticated iterative techniques, with Deep Learning now showing significant promise for further improvements in speed and quality.
  • Cinematic Rendering for Realism: Advanced lighting models like Cinematic Rendering offer a significant leap in realism by simulating light scattering and complex interactions, providing better depth perception and visual fidelity compared to traditional volume rendering.
  • GPU and Parallelism are Key: Both parallel processing and GPU acceleration are crucial for achieving the efficiency needed for large-scale volume rendering and real-time interaction.
  • Digital Twins and SHM: The paper envisions these advancements enabling digital twin models for structural health monitoring, allowing for virtual-physical interaction and real-time defect detection.

Key Results

The paper is primarily a review and survey, so it doesn't present new quantitative results. However, it highlights the following:

  • IR vs. FBP: Iterative Reconstruction (IR) methods generally offer better noise reduction, low-contrast object resolution, and artifact reduction compared to FBP, though often at the cost of slower reconstruction speeds.
  • DLR Potential: Deep Learning Reconstruction (DLR) shows strong potential for noise reduction, low-contrast object resolution, scanning efficiency, and maintaining spatial resolution, with some studies suggesting it can reduce radiation dose by 23% to 76% without compromising image quality.
  • GPU Acceleration Gains: GPU-accelerated methods can achieve efficiency improvements of 6-14 times compared to traditional CPU-based approaches.
  • CR Realism: Cinematic Rendering (CR) provides more natural and realistic visual effects, improving depth perception and recognition, but requires higher computing power, with rendering times ranging from a few seconds to 30 seconds.

Limitations and Future Work

Limitations:

  • DLR Training Data: Deep Learning Reconstruction (DLR) algorithms often rely on high-quality, large-scale training datasets, which can be a bottleneck. The generalization capabilities of trained algorithms can also be limited.
  • CR Computational Cost: Cinematic Rendering (CR) demands substantial computing resources, leading to longer rendering times and making it less suitable for real-time applications compared to traditional methods.
  • Artifacts in DLR: DLR can sometimes generate phantom artifacts that may affect defect judgment.
  • Adaptive Sampling Complexity: Current adaptive sampling algorithms can be complex to design and implement, and may struggle to accurately grasp physical quantity variations.

Future Work:

  • DLR Generalization: Enhancing DLR algorithm generalization and robustness, and exploring dynamic learning capabilities.
  • CR Integration: Integrating CR with VR/AR for comprehensive multi-dimensional visualization and exploring network advancements for remote access and analysis.
  • Hybrid Approaches: Combining AI with optimization algorithms for automatic adjustment of parameters, and exploring hybrid parallel rendering techniques for large-scale datasets.
  • Real-time Interactivity: Addressing high computational costs and long rendering times to meet real-time interactivity requirements, particularly for applications like structural health monitoring.
  • Multi-modal Data Integration: Integrating Micro-CT with other imaging techniques (optical microscopes, electron microscopes) for multi-scale and multi-dimensional information integration.

Summary updated

BibTeX
@article{yin2026three,
  author = {Yipeng Yin and Rao Yao and Qingying Li and Dazhong Wang and Hong Zhou and Zhijun Fang and Jianing Chen and Longjie Qian and Mingyue Wu},
  title = {{Three-dimensional visualization of X-ray micro-CT with large-scale datasets: Efficiency and accuracy for real-time interaction}},
  journal = {Comput. Sci. Rev.},
  year = {2026},
  doi = {10.1016/j.cosrev.2025.100888},
}
Figure