DeepSCI: scalable speckle correlation imaging using physics-enhanced deep learning.

Area of Science:

• Physics
• Computer Science
• Biomedical Imaging

Background:

• Speckle correlation imaging (SCI) is a powerful technique for non-invasive imaging.
• Data-driven approaches for SCI often suffer from limited generalization capabilities.
• Integrating physical principles can enhance the performance and interpretability of deep learning models in imaging.

Purpose of the Study:

• To introduce DeepSCI, a novel physics-enhanced deep learning framework for speckle correlation imaging.
• To demonstrate the ability of DeepSCI to accurately reconstruct images from speckle patterns.
• To address the challenge of limited generalization in data-driven SCI methods.

Main Methods:

• Developed DeepSCI by incorporating the theoretical model of SCI into a neural network architecture.
• Utilized physics priors during both training and testing phases for interpretable data preprocessing and model fine-tuning.
• Evaluated DeepSCI's performance and scalability under medium perturbations and domain shifts.

Main Results:

• DeepSCI accurately reconstructs images from speckle patterns.
• The method exhibits high scalability to perturbations and domain shifts.
• Experimental results confirm DeepSCI's effectiveness in overcoming generalization limitations.

Conclusions:

• DeepSCI offers a robust and effective physics-enhanced deep learning solution for speckle correlation imaging.
• The framework successfully integrates physical knowledge with deep learning for improved imaging performance.
• DeepSCI demonstrates significant advantages over traditional data-driven methods in terms of generalization and accuracy.