CUBDL

CUBDL provides a framework for developing and evaluating deep learning methods for ultrasound image formation.

Key Features:

• International Database: The largest known international database of ultrasound channel data for training and evaluation of deep learning beamforming methods.
• Unified Evaluation Methods: Standardized evaluation protocols that incorporate a neural-network-based global sound speed estimator for consistent model assessment.
• Comprehensive Dataset: A collection of 576 image acquisition sequences, including initially closed evaluation test data from multiple international contributors and additional in vivo breast ultrasound data.
• Computational Artifacts: Includes evaluation code and network weights from challenge winners for reproducible model evaluation and comparison.

Scientific Applications:

• Benchmarking deep-learning beamforming: Provides a standardized platform to compare deep learning beamforming models and to benchmark against delay-and-sum beamforming.
• Image quality improvement: Enables development and evaluation of methods that improve generalized contrast-to-noise ratio (gCNR) and lateral resolution, with challenge networks reporting mean gCNR of 0.81 and mean lateral resolution of 0.32 mm.
• In vivo breast ultrasound research: Supports training and validation of deep learning models on in vivo breast ultrasound data for clinical imaging studies.

Methodology:

Trains and evaluates deep learning models on ultrasound channel data using standardized evaluation protocols and a neural-network-based global sound speed estimator.