Three-Dimensional Thermal Tomography with Physics-Informed Neural Networks.

Area of Science:

• Biomedical Engineering
• Computational Physics
• Artificial Intelligence

Background:

• Accurate internal temperature field reconstruction from surface data is vital for non-invasive thermal imaging, especially for subtle temperature gradients in biological tissues.
• Challenges arise in scenarios with small temperature gradients and non-ideal conditions like noise and background variations.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting internal temperature fields using surface temperature data.
• To assess the model's performance under both ideal and non-ideal conditions, including noise and temperature variations.
• To investigate the impact of physics-informed constraints on model robustness and accuracy.

Main Methods:

• Utilized 3D convolutional neural networks (CNNs) for internal temperature field prediction.
• Incorporated a physics-informed loss function based on the heat equation during training.
• Evaluated model performance on phantoms of varying sizes under simulated noisy and varied background temperature conditions.

Main Results:

• CNNs demonstrated high accuracy for small phantoms but showed reduced predictive capacity in larger domains under non-ideal conditions.
• Physics-informed constraints significantly improved the model's robustness in noisy environments.
• The enhanced model accurately reconstructed deep hot-spots, overcoming limitations of traditional CNNs.

Conclusions:

• Combining deep learning with physical constraints offers a robust framework for high-precision temperature field reconstruction.
• This approach is particularly effective for non-invasive thermal imaging under challenging, non-ideal conditions.
• The study highlights the potential of physics-informed neural networks for advanced thermal sensing applications.