Ultrasound transmission tomography image reconstruction with a fully convolutional neural network

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence

Background:

• Ultrasound computed tomography (UCT) wave-based reconstruction offers superior detail compared to ray-based methods.
• Inverting the wave-based forward model for UCT is computationally intensive.
• Existing methods face challenges in speed and computational demand.

Purpose of the Study:

• To develop an efficient, fully learned image reconstruction method for UCT using convolutional neural networks (CNNs).
• To accelerate the computationally demanding wave-based forward model inversion.
• To improve the speed and robustness of UCT image reconstruction.

Main Methods:

• A CNN-based image reconstruction method was developed, utilizing down-scaling and up-scaling convolutional units.
• A paraxial approximation forward model simulated ultrasound measurement data for training and testing.
• The network was trained on ImageNet data and tested on the OA-Breast Phantom dataset.

Main Results:

• The proposed CNN achieved significantly faster reconstruction times compared to conventional iterative methods (over 20x on CPU, 1000x on GPU).
• The CNN demonstrated comparable image quality to iterative algorithms.
• The method showed increased robustness to noise in image reconstruction.

Conclusions:

• The fully learned CNN approach provides a computationally efficient and effective solution for UCT image reconstruction.
• This method accelerates UCT imaging while maintaining high image quality and noise robustness.
• The developed technique has the potential to enhance diagnostic capabilities in medical imaging.