Model-Based Learning for Accelerated, Limited-View 3-D Photoacoustic Tomography.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Science

Background:

• Deep learning accelerates tomographic reconstructions, enhancing image quality and speed.
• Photoacoustic imaging faces challenges with limited data and artifacts, impacting image resolution.

Purpose of the Study:

• To develop a deep neural network for high-resolution 3-D photoacoustic imaging using restricted measurements.
• To address limited view artifacts in photoacoustic reconstructions through an iterative deep learning approach.

Main Methods:

• A novel deep neural network architecture designed for iterative reconstruction.
• Incorporation of gradient information from data fitting to correct artifacts.
• Separation of training and gradient computation due to forward operator complexity.
• Learning image priors as part of the network training process.

Main Results:

• The network successfully generates high-resolution 3-D images from limited photoacoustic data.
• Demonstrated compensation for limited view artifacts.
• Validated on segmented lung vessels from CT scans and in-vivo photoacoustic data.

Conclusions:

• The proposed deep learning approach enables accurate and high-quality 3-D photoacoustic imaging with improved resolution.
• The method effectively handles data limitations and artifacts inherent in photoacoustic measurements.