Deep learning-based spatio-temporal fusion for high-fidelity ultra-high-speed X-ray radiography

Area of Science:

• Medical Imaging
• Computer Vision
• Materials Science

Background:

• Ultra-high-speed (UHS) X-ray imaging is crucial for dynamic process characterization.
• Exploiting joint acquisition of X-ray videos with distinct configurations remains underexplored.

Purpose of the Study:

• To develop and evaluate a deep learning-based spatio-temporal fusion (STF) framework.
• To reconstruct high spatial resolution, high frame rate, and high fidelity X-ray image sequences.

Main Methods:

• Applied a transfer learning strategy to train the STF model.
• Fused complementary X-ray image sequences with varying resolutions and frame rates.
• Compared STF performance against baseline deep learning, Bayesian fusion, and bicubic interpolation using PSNR, AAD, and SSIM metrics.

Main Results:

• The proposed STF framework significantly outperformed baseline methods in image reconstruction.
• Achieved high average peak signal-to-noise ratios (PSNR) of 37.57 dB and 35.15 dB under specific input conditions.
• Demonstrated robustness across different input frame separations and noise levels.

Conclusions:

• The STF framework effectively enhances the quality of UHS X-ray imaging.
• This approach promises to increase the performance and scientific value of UHS X-ray experiments.
• Integration with high-speed cameras will further amplify its capabilities.