Learning robust features for scatter removal and reconstruction in dynamic ICF X-ray tomography.

Area of Science:

• Medical Imaging
• Computational Physics
• Machine Learning

Background:

• Density reconstruction from X-ray projections is crucial for X-ray computed tomography (CT).
• Noise and scatter in projections significantly degrade reconstruction accuracy.
• Deep learning methods offer potential for improved density reconstruction.

Purpose of the Study:

• To propose a deep learning-based encoder-decoder framework for density reconstruction.
• To evaluate different latent-space feature representations: physics-inspired supervision, self-supervision, and no supervision.
• To assess the performance of the proposed methods under varying noise and scatter conditions.

Main Methods:

• Developed a deep learning encoder-decoder framework for X-ray density reconstruction.
• Explored physics-inspired, self-supervised, and unsupervised latent-space feature representations.
• Compared the proposed deep learning methods against a traditional iterative technique.

Main Results:

• Self-supervised and physics-inspired supervised feature variants demonstrated superior performance with unknown noise and scatter.
• The self-supervised variant achieved the best results in extreme noise scenarios.
• The proposed deep learning methods outperformed a traditional iterative technique in accuracy.

Conclusions:

• Deep learning, particularly with self-supervised or physics-inspired features, enhances density reconstruction accuracy.
• The proposed framework effectively handles noise and scatter in X-ray projections.
• This approach offers a significant improvement over conventional methods for CT applications.