Highly robust reconstruction framework for three-dimensional optical imaging based on physical model constrained...

Area of Science:

• Biomedical Optics
• Computational Imaging
• Machine Learning for Science

Background:

• Three-dimensional optical imaging reconstruction from surface measurements is an ill-posed problem.
• Traditional methods rely heavily on prior information, limiting stability and adaptability.
• Existing deep learning approaches require extensive training data, leading to poor generalization and long acquisition times.

Purpose of the Study:

• To develop a highly robust reconstruction framework for three-dimensional optical imaging.
• To overcome the limitations of traditional regularization-based and data-driven deep learning methods.
• To improve accuracy, stability, and generalizability in optical imaging reconstruction.

Main Methods:

• A physical model-constrained neural network framework was proposed.
• Neural networks generate target distributions from surface measurements.
• A physical model calculates surface light distribution, with mean square error used as a loss function for optimization.
• A movable region strategy was employed to reduce reliance on prior information.

Main Results:

• The framework demonstrated high accuracy, stability, and versatility in simulations and in vivo experiments.
• Performance was robust against different target distributions, noise, and depth variations.
• High spatial resolution was achieved.

Conclusions:

• The proposed framework offers a robust and generalizable solution for 3D optical imaging reconstruction.
• It eliminates the need for manual parameter tuning and training datasets.
• This approach provides a new perspective for optical imaging reconstruction, saving time and resources.