PiLocNet: physics-informed neural network on 3D localization with a rotating point spread function.

Area of Science:

• Optics
• Computational Imaging
• Machine Learning

Background:

• 3D localization is crucial in various imaging applications.
• Existing methods often rely on either model-based optimization or neural networks.
• A gap exists in integrating physical principles into neural network approaches for localization.

Purpose of the Study:

• To introduce PiLocNet, an enhanced physics-informed neural network for 3D localization.
• To combine the strengths of model-based and neural network methods.
• To improve the accuracy and robustness of 3D localization, particularly under noisy conditions.

Main Methods:

• Developed PiLocNet, a physics-informed neural network building upon the previous LocNet.
• Incorporated forward-model-based information via a data-fitting loss term.
• Integrated variational method regularization terms to enhance noise robustness.

Main Results:

• PiLocNet demonstrates superior performance in 3D localization compared to previous methods.
• The physics-informed approach yields physically sensible results.
• The network shows improved robustness against Poisson and Gaussian noise models.

Conclusions:

• PiLocNet effectively combines physical modeling with neural networks for improved 3D localization.
• The framework offers interpretability and enhanced performance.
• The method is adaptable to various point spread functions and imaging problems.