Related Experiment Video
Updated: Jan 8, 2026
06:54
Photorealistic Learned Landscapes for Augmented Reality
Published on: June 27, 2025
641
Scaling deep learning for material imaging with a pseudo 3D model for domain transfer.
Kunning Tang1, Ryan T Armstrong2, Peyman Mostaghimi2
1School of Minerals and Energy Resources Engineering, University of New South Wales, Sydney, Australia.
Nature Communications
|December 12, 2025
Summary
P3T-Net, a novel pseudo-3D domain transfer network, unifies diverse 3D X-ray images for consistent deep learning analysis. This approach enhances model generalizability and reduces retraining needs across various material imaging applications.
Area of Science:
- Materials Science
- Computer Science
- Image Processing
Background:
- Deep learning significantly advanced 3D X-ray imaging for material characterization.
- Variations in imaging conditions cause deep learning model inconsistencies, requiring frequent retraining.
- Lack of domain consistency limits the generalizability and applicability of deep learning models in material imaging.
Purpose of the Study:
- To introduce P3T-Net, a pseudo-3D domain transfer network, for unifying diverse 3D image datasets.
- To enable deep learning models to process multiple datasets without retraining.
- To reduce computational costs associated with 3D image domain transfer.
Main Methods:
- Developed P3T-Net, a pseudo-3D domain transfer network.
- Transferred diverse 3D images into a uniform domain for deep learning processing.
- Demonstrated P3T-Net on geological rock, hydrogen fuel cells, and lithium-ion battery imaging.
Main Results:
- P3T-Net enables the reuse of previously trained deep learning networks for new image datasets.
- Achieved image enhancement for fast scans and multi-source imaging.
- Enabled accurate segmentation across different imaging conditions.
- Demonstrated tera-scale 3D image transfer (10^11 voxels) on a single GPU.
Conclusions:
- P3T-Net effectively addresses cross-domain inconsistencies in material imaging.
- The proposed approach enhances the robustness and generalizability of deep learning solutions.
- P3T-Net offers a computationally efficient method for 3D image domain transfer.
