Medical Microwave Imaging Using Physics-Guided Deep Learning-Part 1: The Forward Solver.

Area of Science:

• Computational electromagnetics
• Deep learning applications
• Physics-informed machine learning

Background:

• Deep neural networks offer faster solutions for electromagnetic inverse problems but require physics frameworks for accuracy.
• Current physics-guided deep learning solvers face training bottlenecks due to reliance on computationally intensive forward solvers.
• Ensuring physically correct results necessitates robust forward solvers within the deep learning training loop.

Purpose of the Study:

• To develop a fast and accurate self-supervised deep learning forward solver for electromagnetic problems.
• To overcome the efficiency limitations of existing physics-guided deep learning inverse solvers.
• To enable reliable and practical deep learning-based solutions for inverse electromagnetic problems.

Main Methods:

• A physics-based framework dividing the domain into interior (scatterers) and exterior (background) regions.
• A hybrid loss function combining Maxwell's curl equation and integral equation with Green's function.
• Self-supervised learning guiding the neural network to generate accurate scattered fields.

Main Results:

• The solver achieves high global and local accuracy, verified on random and realistic models.
• Over 95% of test cases show <0.15 root-mean-square error in scattered field and dielectric properties.
• The method demonstrates a 97% speedup compared to conventional solvers, outperforming recent deep learning approaches.

Conclusions:

• The presented self-supervised deep learning forward solver is accurate, efficient, and generalizable.
• This approach effectively addresses the training bottleneck in physics-guided deep learning for inverse problems.
• The developed solver facilitates the creation of more reliable and practical deep learning-based inverse solvers.