Evaluating Input Domain and Model Selection for Deep Network Ultrasound Beamforming.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Ultrasound
• Signal Processing

Background:

• Ultrasound B-mode image quality enhancement is crucial.
• Deep neural networks (DNNs) show promise for efficient beamforming.
• Previous work focused on frequency-domain DNNs, but direct time-domain comparisons are lacking.

Purpose of the Study:

• To systematically compare frequency- and time-domain DNN implementations for ultrasound beamforming.
• To introduce contrast-to-noise ratio (CNR)-based regularization for improved model selection.
• To evaluate robustness to varying pulse shapes and assess performance on diverse datasets.

Main Methods:

• Generated training and testing channel data using simulations (anechoic cysts) and physical phantoms.
• Implemented and compared frequency- and time-domain DNNs for beamforming.
• Applied CNR-based regularization during DNN training.

Main Results:

• Simplified time-domain DNNs demonstrated robustness, particularly with phase-preserving data.
• Both frequency- and time-domain DNNs achieved median in vivo CNR improvements of 0.39 dB and 0.36 dB over conventional DAS, respectively.
• CNR regularization significantly improved correlations between training loss and simulated/in vivo CNR.

Conclusions:

• Time-domain DNNs are a viable and robust alternative for ultrasound beamforming.
• CNR-based regularization enhances DNN model selection accuracy and clinical relevance.
• These advancements offer more efficient and effective ultrasound image quality improvement.