Deep-learning projector for optical diffraction tomography.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Inverse Problems

Background:

• Optical diffraction tomography (ODT) is crucial for estimating refractive indices of unknown objects.
• ODT involves solving ill-posed inverse problems governed by the wave equation.
• Traditional solutions minimize an objective function with data-fidelity and regularization terms.

Purpose of the Study:

• To develop a novel method for optical diffraction tomography using deep learning.
• To integrate convolutional neural networks (CNNs) into projected-gradient-descent algorithms.
• To improve the accuracy and robustness of refractive index estimation.

Main Methods:

• Training a CNN as a projector within a projected-gradient-descent framework.
• Iterative refinement of estimates to ensure measurement consistency and high quality.
• Utilizing a hybrid approach combining CNN-based and regularization-based techniques.

Main Results:

• Demonstrated improved performance on 2D simulated and real data compared to conventional and deep learning methods.
• Achieved high-quality refractive index estimates with enhanced measurement consistency.
• Validated the generalizability of the trained CNN projector for various forward models.

Conclusions:

• The proposed CNN-based projector offers a significant advancement in optical diffraction tomography.
• This hybrid approach effectively balances data fidelity with prior constraints for superior results.
• The method shows promise for handling complex scattering events and diverse applications.