Noise-resilient deep learning for integrated circuit tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Materials Science

Background:

• X-ray tomography is a crucial non-destructive imaging method.
• High-fidelity reconstruction often requires regularization priors, especially with sparse-view or low-photon data.
• Deep learning models have shown promise in X-ray tomography by learning priors from training data.

Purpose of the Study:

• To develop a noise-resilient deep-reconstruction algorithm for X-ray tomography.
• To improve reconstruction quality under challenging imaging conditions, such as low photon counts.
• To apply the algorithm to integrated circuit tomography.

Main Methods:

• A novel deep-reconstruction algorithm was proposed.
• The neural network was trained using regularized reconstructions from a conventional algorithm.
• The approach focuses on learning noise resilience directly within the network's prior.

Main Results:

• The proposed algorithm demonstrated strong noise resilience without requiring explicit training on noisy data.
• Acceptable reconstructions were achieved with significantly fewer photons in test data.
• The method proved effective for integrated circuit tomography.

Conclusions:

• The developed noise-resilient deep-reconstruction framework enhances X-ray tomography performance under low-photon sampling.
• This approach offers a pathway to high-quality tomographic imaging even when acquisition times are limited.
• The learned prior exhibits robustness to variations in noise characteristics.