Model-informed deep-learning photoacoustic reconstruction for low-element linear array.

Area of Science:

• Biomedical Imaging
• Medical Physics
• Deep Learning

Background:

• Photoacoustic tomography (PAT) using linear array ultrasound transducers faces image quality challenges due to sparse sensor arrangements and limited pitch.
• Conventional reconstruction methods like delay-and-sum and model-based techniques are either inaccurate or computationally intensive.
• Wearable PAT systems require reduced complexity and power consumption, exacerbating these limitations.

Purpose of the Study:

• To develop a simplified and efficient deep learning framework for PAT image reconstruction tailored to linear array transducers.
• To address the computational challenges of model matrix inversion in model-based deep learning for PAT.
• To improve image reconstruction quality and reduce computational load for PAT systems, especially those with fewer sensors.

Main Methods:

• Introduction of a lightweight Generative-Equivalent Convolutional Neural Network (GE-CNN) framework.
• Optimization of the model matrix size and adjoint transformations for reduced computational demand.
• Rigorous evaluation using synthetic data, experimental phantoms, and in-vivo rat liver imaging.

Main Results:

• Achieved a 4-fold reduction in model matrix size (e.g., 2.09 GB for 32 elements vs. 8.38 GB for 128 elements).
• Accelerated processing by approximately 46.3%, reducing reconstruction time from 7.88 to 4.23 seconds.
• Demonstrated improved reconstruction performance with minimal hardware requirements.

Conclusions:

• The proposed lightweight GE-CNN framework offers an efficient solution for PAT image reconstruction with linear array transducers.
• This approach significantly enhances reconstruction quality while substantially reducing computational complexity and processing time.
• The method shows promise for practical implementation in resource-constrained and wearable PAT devices.