Efficient and accurate inversion of multiple scattering with deep learning.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning Applications

Background:

• Multiple light scattering poses a significant challenge in imaging applications like diffraction tomography.
• Current image reconstruction methods often rely on complex, non-convex optimization, which can be slow and computationally intensive.
• Accurate refractive index imaging is critical for understanding material properties and biological structures.

Purpose of the Study:

• To develop a novel, efficient, and high-performance image reconstruction method for multiple light scattering scenarios.
• To leverage deep convolutional neural networks (CNNs) as an alternative to traditional optimization-based approaches.
• To demonstrate the superiority of the proposed CNN-based method in terms of speed and image quality.

Main Methods:

• Design and training of a deep convolutional neural network (CNN) specifically for inverting multiple scattered light measurements.
• Utilizing a nonlinear measurement model to accurately represent the physics of multiple scattering.
• Validation using both simulated and experimental datasets to assess reconstruction accuracy and speed.

Main Results:

• The proposed CNN-based method achieved substantially faster image reconstruction compared to state-of-the-art optimization techniques.
• The deep learning approach yielded higher imaging quality, producing more accurate refractive index maps.
• The method proved effective on both simulated data and real-world experimental measurements.

Conclusions:

• Deep convolutional neural networks offer a powerful and efficient alternative for image reconstruction in the presence of multiple light scattering.
• The developed CNN approach significantly advances the state-of-the-art in diffraction tomography and related imaging fields.
• This work paves the way for faster and more accurate non-invasive imaging in various scientific and industrial applications.