PnP Restoration with Domain Adaptation for SANS

Shirin Shoushtari, Edward P. Chandler, Jialiang Zhang, Manjula Senanayake, Sai Venkatesh Pingali, Marcus Foston, Ulugbek S. KamilovarXiv 2024 · DOI · arXiv

TL;DR

This paper introduces PR-SANS, a novel deep learning-based image restoration method for Small Angle Neutron Scattering (SANS) data. By employing domain adaptation to fine-tune generic image priors on limited SANS data, PR-SANS significantly enhances the signal-to-noise ratio (SNR) of low-quality SANS images, enabling a 12x reduction in acquisition time without sacrificing data quality.

What Problem Does This Paper Address?

Small Angle Neutron Scattering (SANS) is a crucial technique for understanding material structures at the nanoscale. However, obtaining detailed structural information often requires long acquisition times, which limits its application in dynamic, in-situ studies. While accelerating data acquisition is desirable, it typically leads to reduced signal-to-noise ratios (SNR) in the resulting SANS detector images. Deep learning (DL) methods have shown promise in improving image quality, but training effective DL models for SANS requires substantial amounts of experimental data, which is often scarce and expensive to acquire. This paper tackles the challenge of restoring high-SNR SANS images from low-SNR measurements when limited experimental SANS data is available for training DL models.

Methods

The core of the proposed method, PR-SANS, is a Plug-and-Play (PnP) algorithm. PnP methods leverage powerful DL denoisers as "priors" within an iterative optimization framework for image restoration. The novelty here lies in the domain adaptation strategy for these DL priors.

  1. Pre-training: A generic image denoiser (specifically, a DnCNN model) is first trained on a large dataset of natural grayscale images to learn general image features.
  2. Domain Adaptation: This pre-trained denoiser is then fine-tuned using a limited number of experimental SANS detector image pairs (low-SNR and high-SNR). This adaptation process tailors the generic prior to the specific characteristics of SANS data.
  3. PR-SANS Algorithm: The adapted denoiser is then plugged into a PnP framework. This framework iteratively refines an estimate of the clean SANS image by alternating between a data-fidelity step (ensuring consistency with the noisy measurement) and a denoising step (using the adapted DL prior).

The authors also provide a theoretical convergence analysis of PR-SANS, specifically addressing the impact of using inexact, domain-adapted priors instead of ideal ones.

Datasets

  • Source Domain (for pre-training): A dataset of natural grayscale images (400 pairs mentioned in the text, with a specific reference to [45] for a database of human segmented natural images).
  • Target Domain (for adaptation and evaluation): Experimental SANS detector images. The paper uses 100 pairs of low- and high-SNR SANS detector images for training the adapted priors and a separate set of 10 pairs for testing.

Key Insights and Ideas

  • Domain Adaptation is Crucial for Scarce Data: The paper strongly demonstrates that adapting a generic DL prior to the specific domain of SANS data is significantly more effective than using a "zero-start" (untrained or minimally trained) prior, especially when experimental SANS data is limited.
  • PnP Framework Leverages Adapted Priors Effectively: The PnP approach provides a robust framework to integrate these domain-adapted DL priors, allowing for high-quality image restoration even from very noisy inputs.
  • Theoretical Guarantees for Adapted Priors: The theoretical analysis provides a foundation for understanding the convergence properties of PR-SANS, showing that while adapted priors introduce an error, the algorithm still asymptotically approaches a stationary point of the objective function.
  • Sample Complexity Matters: The performance of the adapted prior improves as more SANS data is used for fine-tuning, highlighting the trade-off between data availability and restoration quality.

Key Results

  • Significant SNR Improvement: PR-SANS with adapted priors can recover high-SNR 2D SANS detector images from low-SNR measurements. For instance, one result shows an increase in SNR from 18.44 dB (Low-SNR data) to 22.36 dB (PR-SANS Adapted).
  • 12x Faster Acquisition: The method enables a reduction in acquisition times by a factor of 12 while maintaining comparable signal quality to the original, longer acquisitions.
  • Superior Performance: PR-SANS with adapted priors outperforms other methods, including Total Variation (TV) denoising and end-to-end DnCNN models (both zero-start and adapted), in terms of SNR and RMSE for 2D detector images and 1D intensity plots.
  • Visual Quality: Figures 3 and 5 show compelling visual improvements, with PR-SANS producing sharper and cleaner images with more discernible features compared to baseline methods.

Limitations and Future Work

  • Sample Complexity: While domain adaptation is effective, the performance still improves with more SANS data for adaptation. The paper suggests that around 40-100 SANS data pairs are adequate for obtaining good adapted priors.
  • Ideal Prior Unavailability: The theoretical analysis acknowledges that obtaining the "ideal" SANS prior is practically infeasible, and the error bound depends on the discrepancy between the ideal and adapted priors.
  • Future Work: The authors suggest further investigation into the relationship between sample complexity and restoration performance, and potentially exploring other domain adaptation strategies. The paper also implicitly points towards the potential for further acceleration of SANS experiments with such advanced restoration techniques.

Summary updated

Abstract

Small Angle Neutron Scattering (SANS) is a non-destructive technique utilized to probe the nano- to mesoscale structure of materials by analyzing the scattering pattern of neutrons. Accelerating SANS acquisition for in-situ analysis is essential, but it often reduces the signal-to-noise ratio (SNR), highlighting the need for methods to enhance SNR even with short acquisition times. While deep learning (DL) can be used for enhancing SNR of low quality SANS, the amount of experimental data available for training is usually severely limited. We address this issue by proposing a Plug-and-play Restoration for SANS (PR-SANS) that uses domain-adapted priors. The prior in PR-SANS is initially trained on a set of generic images and subsequently fine-tuned using a limited amount of experimental SANS data. We present a theoretical convergence analysis of PR-SANS by focusing on the error resulting from using inexact domain-adapted priors instead of the ideal ones. We demonstrate with experimentally collected SANS data that PR-SANS can recover high-SNR 2D SANS detector images from low-SNR detector images, effectively increasing the SNR. This advancement enables a reduction in acquisition times by a factor of 12 while maintaining the original signal quality.
BibTeX
@misc{shoushtari2024pnp,
  author = {Shirin Shoushtari and Edward P. Chandler and Jialiang Zhang and Manjula Senanayake and Sai Venkatesh Pingali and Marcus Foston and Ulugbek S. Kamilov},
  title = {{PnP Restoration with Domain Adaptation for SANS}},
  year = {2024},
  eprint = {2403.10495},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2403.10495},
}
Figure