Learning Robust Features for Scatter Removal and Reconstruction in Dynamic ICF X-Ray Tomography

Density reconstruction from X-ray projections is an important problem in radiography with key applications in scientific and industrial X-ray computed tomography (CT). Often, such projections are corrupted by unknown sources of noise and scatter, which when not properly accounted for, can lead to significant errors in density reconstruction. In the setting of this problem, recent deep learning-based methods have shown promise in improving the accuracy of density reconstruction. In this article, we propose a deep learning-based encoder-decoder framework wherein the encoder extracts robust features from noisy/corrupted X-ray projections and the decoder reconstructs the density field from the features extracted by the encoder. We explore three options for the latent-space representation of features: physics-inspired supervision, self-supervision, and no supervision. We find that variants based on self-supervised and physicsinspired supervised features perform better over a range of unknown scatter and noise. In extreme noise settings, the variant with self-supervised features performs best. After investigating further details of the proposed deep-learning methods, we conclude by demonstrating that the newly proposed methods are able to achieve higher accuracy in density reconstruction when compared to a traditional iterative technique.

💡 Research Summary

This paper addresses the critical challenge of accurately reconstructing the 3D density field of an object from its X-ray projection data, which is often severely corrupted by unknown scatter and noise, particularly in dynamic settings like Inertial Confinement Fusion (ICF) experiments. Traditional methods like Filtered Back Projection (FBP) or Model-Based Iterative Reconstruction (MBIR) rely heavily on accurate forward and noise models, which are difficult to obtain in practice, leading to significant reconstruction errors.

The authors propose a novel deep learning framework centered on learning “robust features” from corrupted projections. The core architecture is an encoder-decoder network. The encoder takes noisy X-ray transmission images (T) as input and is tasked with extracting a latent feature representation that is invariant to the corrupting scatter and noise. The decoder then maps these robust features back to the desired clean density volume (ρ).

A key contribution is the realistic simulation pipeline for generating training data. Starting from hydrodynamic simulations (solving the Euler equations) of a double-shell ICF capsule, clean synthetic radiographs (D) are generated using a monoenergetic forward model. These are then corrupted by a comprehensive perturbation model that includes source/detector blur (2D Gaussian kernels), correlated scatter (another Gaussian kernel), a polynomial background scatter field, and Poisson-distributed gamma and photon noise. This creates a realistic dataset of noisy-clean pairs (T, ρ) for training.

The study innovatively explores and compares three distinct strategies for learning the latent feature representation within the encoder-decoder framework:

1. Physics-Inspired Supervised Latent Representation (PISLR): The encoder is additionally supervised to produce features that match predefined physics-based features—specifically, the shock edge maps extracted (using a Canny filter) from the clean density volumes. This injects prior physical knowledge into the learning process.
2. Self-Supervised Latent Representation (SSLR): The encoder is encouraged to produce similar feature representations for different noisy realizations of the same underlying density. This is achieved via a regularization loss that minimizes the distance between latent features from differently perturbed versions of the same clean projection. This forces the network to learn features that are consistent across noise variations.
3. Unsupervised Latent Representation (ULR): A baseline approach where the latent features are learned with no explicit supervision or regularization, guided only by the final density reconstruction loss.

The numerical experiments demonstrate that both the PISLR and SSLR variants significantly outperform the ULR baseline across a wide range of unknown scatter and noise levels. The most significant finding is that in extreme noise settings, unseen during training, the SSLR method achieves the best reconstruction accuracy (lowest relative L2 error). This indicates that the self-supervised learning of noise-invariant features provides superior generalization and robustness compared to relying on predefined physical features (PISLR), which may not capture all relevant information or may be less adaptable to severe corruption.

In conclusion, the paper successfully demonstrates that learning robust, noise-invariant features via deep learning, particularly through a self-supervised paradigm, is a powerful approach for density reconstruction in severely corrupted X-ray tomography. The proposed methods achieve higher accuracy than a traditional iterative technique, offering a promising data-driven solution for challenging imaging problems in scientific domains like ICF where accurate physical models of corruption are elusive.