GCI: A generative CT interpolation model for rapid wood CT scanning

Area of Science:

• Biomaterials imaging
• Computational imaging
• Wood science

Background:

• Micro-computed tomography (micro-CT) provides 3D imaging but faces limitations in scan time, cost, and data acquisition speed.
• Fast scanning strategies in micro-CT can lead to information loss, impacting image quality and detailed analysis.
• Existing interpolation methods struggle with complex structural variations in wood, causing artifacts and blurring.

Purpose of the Study:

• To develop a generative model for reconstructing high-quality 3D wood images from fast micro-CT scans.
• To enable efficient and accurate 3D characterization of wood and other biomass materials.
• To overcome the limitations of traditional interpolation methods in preserving details and structure.

Main Methods:

• Proposed a generative CT interpolation (GCI) model utilizing a 3D ResNet-18 backbone and a 3D CBAM attention mechanism.
• Employed an encoder-decoder architecture for multi-scale feature fusion to enhance image reconstruction.
• Constructed the CT-WOOD dataset, the first dedicated dataset for wood CT Z-axis super-resolution.

Main Results:

• GCI significantly outperforms existing interpolation methods in key metrics like PSNR, SSIM, and LPIPS, with a 22% improvement in perceptual similarity.
• Achieved up to a fourfold reduction in scanning time while maintaining high image quality.
• Demonstrated superior detail restoration and structural preservation compared to traditional methods through visual and segmentation experiments.

Conclusions:

• GCI offers a reliable technical solution for fast and high-quality 3D imaging of wood using micro-CT.
• The model effectively recovers information lost during fast scanning, enabling efficient material characterization.
• This approach paves the way for large-scale application of micro-CT in wood science and biomass research.