Multiple-scattering simulator-trained neural network for intensity diffraction tomography.

Area of Science:

• Biophysics
• Computational Biology
• Microscopy

Background:

• Traditional 3D phase recovery methods for biological samples are computationally intensive.
• Achieving high physical model accuracy often compromises processing speed and efficiency.

Purpose of the Study:

• To develop a computationally efficient deep learning framework for high-speed 3D phase recovery.
• To reconstruct complex biological samples with improved accuracy and speed using intensity diffraction tomography.

Main Methods:

• Implemented an approximant-guided deep learning framework with a lightweight 2D network structure.
• Utilized a multi-channel input for axial information encoding and a physics model simulator-based learning strategy.
• Trained the network on natural image datasets and validated on biological samples like cells and C. elegans.

Main Results:

• Successfully reconstructed complex 3D biological samples, including weakly and strongly scattering specimens.
• Demonstrated robustness by reconstructing dynamic samples from live C. elegans videos.
• Showcased generalization capabilities on algae samples from varied experimental setups.

Conclusions:

• The deep learning framework offers a significant improvement in computational efficiency for 3D phase recovery.
• The method achieves high reconstruction quality and robustness, outperforming traditional algorithms.
• This approach advances high-speed imaging of dynamic and complex biological structures.