Posterior Sampling with Latent Diffusion for Microwave Brain Imaging.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Computational Electromagnetics

Background:

• Microwave imaging offers a non-invasive approach for stroke detection and monitoring.
• Stroke detection via microwave imaging is an ill-posed problem requiring effective prior information.
• Conventional methods with hand-crafted priors limit image resolution.

Purpose of the Study:

• To develop a physics-data-driven method for high-resolution microwave imaging of the head.
• To improve the accuracy of detecting and monitoring strokes using microwave imaging.
• To overcome limitations of classical methods in exploiting prior knowledge for stroke imaging.

Main Methods:

• A physics-data-driven approach sampling from the posterior distribution of head electrical properties.
• Integration of a data consistency sampler for measurement alignment.
• Utilization of a latent diffusion model and variational auto-encoder as an expressive prior.

Main Results:

• The proposed method achieved superior image resolution for electrical property distribution.
• Demonstrated avoidance of local minima common in deterministic inversion techniques.
• Outperformed existing methods with a lower model misfit (0.067 vs. ≥ 0.085) and higher SSIM (0.936 vs. ≤ 0.906).

Conclusions:

• The developed method significantly enhances microwave imaging for stroke detection.
• This approach offers a promising solution for accurate and high-resolution stroke monitoring.
• The physics-data-driven strategy effectively leverages prior information for improved imaging outcomes.