Physics-driven learning of Wasserstein GAN for density reconstruction in dynamic tomography

Area of Science:

• Medical Imaging
• Computational Physics
• Machine Learning

Background:

• Accurate object density reconstruction is crucial but challenged by scattered radiation and noise.
• Traditional methods struggle with unmodeled scatter and artifacts, limiting accuracy in dynamic imaging.
• Machine learning offers potential for improved density reconstruction, especially for time-evolving fields.

Purpose of the Study:

• To demonstrate the effectiveness of deep neural networks for artifact removal in noisy density reconstructions.
• To develop a machine learning model capable of handling imperfectly characterized noise.
• To enhance the accuracy of density reconstruction in dynamic imaging applications.

Main Methods:

• Utilized a Wasserstein generative adversarial network (WGAN) where the generator acts as a denoiser.
• Trained networks on large density time-series datasets with simulated noise distributions.
• Incorporated a supervised loss and physics-based constraints (e.g., mass conservation) during training.

Main Results:

• The WGAN effectively removed significant portions of unknown noise from density time-series data.
• The inclusion of supervised loss improved density restoration performance.
• Physics-based constraints further enabled highly accurate density reconstructions.

Conclusions:

• Learned deep neural networks, specifically WGANs, show strong potential for artifact removal in noisy density reconstructions.
• The developed framework successfully handles imperfectly characterized noise and experimental artifacts.
• This approach advances accurate density reconstruction, particularly for dynamic imaging scenarios.