Hardware-Independent Deep Signal Processing: A Feasibility Study in Echocardiography

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ultrasound Technology

Background:

• Conventional ultrasound (US) signal processing is complex and hardware-dependent.
• Deep learning (DL) offers a potential alternative for signal processing, promising reduced inference time and hardware portability.
• Existing DL models for US often focus on specific tasks rather than replicating entire processing chains.

Purpose of the Study:

• To develop and evaluate a DL model that replicates the BMode signal processing chain of a high-end US system.
• To assess the model's performance with different probes and on lower-end US systems.
• To explore the potential for transferring advanced US features to less sophisticated hardware.

Main Methods:

• A deep neural network (DNN) was trained using supervised learning to map raw in-phase and quadrature data to processed US images.
• The training dataset comprised 30,000 cardiac image frames from a GE HealthCare Vivid E95 system.
• The DL model replicated key processing steps including filtering, compounding, and compression.

Main Results:

• The lightweight DL model accurately replicated the commercial scanner's signal processing chain.
• A structural similarity index measure (SSIM) of 98.56 ± 0.49 was achieved on a test dataset.
• The DL model demonstrated equivalent or improved image quality when applied to data from a different probe.
• Enhanced image quality was observed when applied to a Verasonics dataset, indicating potential for porting features to lower-end systems.

Conclusions:

• A single DL model can effectively replicate a high-end US system's BMode processing chain for specific applications.
• DL models show promise for cost-effective tuning and implementation strategies for US vendors.
• The developed DL model facilitates the transfer of advanced US imaging capabilities to lower-end systems, enhancing accessibility and performance.