CohereNet: A Deep Learning Architecture for Ultrasound Spatial Correlation Estimation and Coherence-Based Beamforming

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep neural networks (DNNs) are powerful universal approximators capable of learning complex functions.
• Spatial coherence functions are crucial for coherence-based beamforming techniques, such as short-lag spatial coherence (SLSC) beamforming.
• Current SLSC methods can be computationally intensive, limiting their application in certain scenarios.

Purpose of the Study:

• To utilize DNNs to estimate spatial coherence functions for improved coherence-based beamforming.
• To develop and evaluate CohereNet, a custom DNN for estimating spatial coherence functions.
• To assess the performance and computational efficiency of CohereNet compared to traditional methods.

Main Methods:

• A fully connected DNN (CohereNet) was designed and trained to estimate spatial coherence functions from ultrasound data.
• The DNN was trained on in vivo breast data from 18 patients.
• Performance was evaluated on independent datasets including in vivo breast, liver, and phantom data using various transducer geometries and ultrasound systems.

Main Results:

• CohereNet achieved a mean correlation of 0.93 with CPU-computed SLSC images.
• The DNN-based SLSC approach was up to 3.4 times faster than CPU-based methods.
• CohereNet demonstrated improved image quality and more consistent computational times compared to GPU-based approaches.

Conclusions:

• Deep learning offers a promising alternative for estimating correlation functions from ultrasound data.
• CohereNet shows potential for applications in various ultrasound imaging and beamforming areas, including speckle tracking, elastography, and blood flow estimation.
• The DNN approach may replace GPU-based methods in low-power, remote, and synchronization-dependent applications.