Efficient Photoacoustic Image Synthesis with Deep Learning.

Area of Science:

• Biomedical Optics
• Medical Imaging
• Computational Biology

Background:

• Photoacoustic imaging offers real-time visualization of functional tissue parameters like oxygenation.
• Quantifying these parameters in photoacoustic imaging is challenging.
• Deep learning (DL) shows promise for solving this problem, but efficient training and validation methods are lacking.

Purpose of the Study:

• To investigate the use of DL for accurate and efficient simulation of photon propagation in biological tissue.
• To enable photoacoustic image synthesis using DL models.
• To develop back-propagatable neural networks for improved quantification in photoacoustic imaging.

Main Methods:

• Developed a DL approach using neural networks (U-Net and Fourier Neural Operator) to estimate initial pressure distribution from optical properties.
• Trained models on synthetic data for simulating photon propagation.
• Validated performance using in silico multispectral human forearm images, comparing against Monte Carlo simulations.

Main Results:

• DL methods achieved a 100x speedup in image generation compared to standard Monte Carlo simulations.
• Both U-Net and Fourier Neural Operator (FNO) models demonstrated high accuracy, with FNO slightly outperforming U-Net.
• DL models function as differentiable surrogate models, enabling gradient-based optimization in the photoacoustic image synthesis pipeline.

Conclusions:

• DL models can accurately and efficiently simulate photon propagation for photoacoustic image synthesis.
• The efficiency of DL methods allows for large-scale training data generation, potentially expediting clinical translation of photoacoustic imaging.
• Differentiable DL models offer advantages for error back-propagation and optimization in photoacoustic imaging workflows.