Comparing Deep Learning Frameworks for Photoacoustic Tomography Image Reconstruction.

Area of Science:

• Medical Imaging
• Computational Imaging
• Biomedical Engineering

Background:

• Conventional photoacoustic image reconstruction methods struggle with sparse sensing and geometric constraints.
• Learning-based approaches are emerging to enhance photoacoustic tomography (PAT) reconstruction quality.

Purpose of the Study:

• To systematically compare and evaluate recent learning-based methods and modified networks for PAT image reconstruction.
• To investigate both learning-based post-processing and model-based learned iterative reconstruction methods.
• To assess the impact of input data characteristics on reconstruction performance.

Main Methods:

• Investigated learning-based post-processing methods for photoacoustic image reconstruction.
• Evaluated model-based learned iterative reconstruction methods.
• Analyzed the influence of different input data on reconstruction outcomes.

Main Results:

• Reconstruction performance is primarily driven by the information content of the input data.
• Minimal performance differences were observed among various learning-based post-processing models.
• Model-based learned iterative reconstruction demonstrated superior generalizability and robustness.

Conclusions:

• Input data quality is crucial for effective photoacoustic image reconstruction.
• Model-based learned iterative reconstruction is a more robust and generalizable approach for challenging PAT scenarios.
• Further research should focus on optimizing input data for learning-based reconstruction techniques.