Deep Learning for Ultrasound Image Formation: CUBDL Evaluation Framework and Open Datasets

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ultrasound Technology

Background:

• Deep learning shows promise for advancing ultrasound image formation, balancing image quality and speed.
• A lack of standardized datasets and evaluation methods hinders optimal solution development.

Purpose of the Study:

• Introduce the largest international database of ultrasound channel data.
• Describe unified evaluation methods for deep learning in ultrasound beamforming.
• Facilitate reproducible research and development in the field.

Main Methods:

• Compiled a database of 576 ultrasound image acquisition sequences.
• Developed and applied evaluation methods from the Challenge on Ultrasound Beamforming with Deep Learning (CUBDL).
• Utilized a neural-network-based global sound speed estimator for fair evaluation.

Main Results:

• Deep learning networks achieved superior performance, with a mean gCNR of 0.81 and lateral resolution of 0.32 mm.
• Compared to delay-and-sum beamforming (gCNR 0.67, resolution 0.42 mm).
• Publicly released datasets, evaluation code, and winning network weights.

Conclusions:

• The CUBDL database and methods provide a foundation for advancing deep learning in ultrasound.
• Standardized resources enable fair comparison and accelerate innovation in ultrasound image formation.
• Public availability promotes collaboration and further research.