Learned regularization for image reconstruction in sparse-view photoacoustic tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Biomedical Engineering

Background:

• Sparse-view measurements in photoacoustic computed tomography (PACT) accelerate data acquisition but challenge image reconstruction.
• Iterative reconstruction with regularization is common but can introduce image artifacts.

Purpose of the Study:

• To develop a learned regularization method for suppressing image artifacts in model-based iterative reconstruction for sparse-view PACT.
• To improve image quality and reconstruction efficiency in PACT under data-constrained acquisition scenarios.

Main Methods:

• A lightweight dual-path neural network was designed to learn regularization features from both data and image domains.
• The method integrates a physical model of photoacoustic imaging with deep learning.
• The network was trained and tested on simulated and in vivo datasets, compared against Tikhonov, total variation, and U-Net based methods.

Main Results:

• The proposed learned regularization method significantly suppressed image artifacts compared to conventional techniques.
• The lightweight network, only 0.15% the size of a U-Net, achieved superior performance.
• Reconstruction converged within five iterations, taking less than one-third the time of traditional methods.

Conclusions:

• Learned regularization offers a powerful approach to enhance sparse-view PACT image reconstruction.
• Integrating deep learning with physical models improves imaging performance in practical PACT applications.
• The method demonstrates potential for faster and higher-quality PACT imaging, even with limited data.