PA OmniNet: A retraining-free, generalizable deep learning framework for robust photoacoustic image reconstruction

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Clinical translation of photoacoustic imaging (PAI) requires cost-effective systems.
• Sparse sampling in PAI reduces hardware costs but introduces reconstruction artifacts, degrading image quality.
• Deep learning models like U-net often need retraining for new system configurations, increasing data and computational demands.

Purpose of the Study:

• To introduce PA OmniNet, a modified U-net model designed for generalized artifact removal in sparse sampling PAI.
• To enable adaptation to new system configurations using minimal example data (context set) without retraining.
• To improve the efficiency and applicability of deep learning in cost-effective PAI systems.

Main Methods:

• Developed PA OmniNet, a U-net variant capable of adapting to different PAI system configurations.
• Utilized a small context set (4-32 images) to condition the model for artifact removal.
• Evaluated PA OmniNet against standard U-net on diverse datasets including in vivo (mouse, human), synthetic, and multi-wavelength data.

Main Results:

• PA OmniNet demonstrated superior generalization across different system configurations compared to standard U-net.
• Achieved average improvements: 8.3% in Structural Similarity Index, 11.6% reduction in Root Mean Square Error, 1.55 dB increase in Peak Signal-to-Noise Ratio.
• In 66% of cases, generalized PA OmniNet outperformed specifically trained U-net models.

Conclusions:

• PA OmniNet effectively removes artifacts in sparse sampling PAI, generalizing across system configurations without retraining.
• The model's adaptability using a small context set significantly enhances its practical utility for clinical translation.
• PA OmniNet offers a promising solution for developing cost-effective and high-quality PAI systems.