Domain Transform Network for Photoacoustic Tomography from Limited-view and Sparsely Sampled Data

Area of Science:

• Medical Imaging
• Computational Imaging
• Biomedical Engineering

Background:

• Deep learning shows promise for photoacoustic tomography (PAT) reconstruction, especially with limited or sparse data.
• Existing deep learning PAT methods are often limited by reliance on conventional linear reconstruction for signal-to-image transformation.

Purpose of the Study:

• To develop a novel deep learning reconstruction approach for PAT that overcomes limitations of conventional methods.
• To enhance image reconstruction quality and speed in PAT using limited-view and sparse data.

Main Methods:

• Proposed a Feature Projection Network (FPnet) for data-driven signal-to-image transformation in PAT.
• Integrated an image post-processing U-net to further refine reconstruction results.
• Employed specific data pre-processing and training strategies tailored for FPnet.

Main Results:

• Achieved high-quality image reconstruction from limited-view and sparse PAT data.
• Demonstrated superior performance compared to methods relying on linear reconstruction.
• Attained a reconstruction speed of 15 frames per second with GPU acceleration.

Conclusions:

• The proposed FPnet approach offers a powerful, data-driven alternative for PAT image reconstruction.
• This method significantly improves reconstruction quality and efficiency for limited and sparse data scenarios.
• The integration of FPnet and U-net presents a promising advancement in medical image reconstruction technology.