Three-Dimensional Convolutional Neural Network for Ultrasound Surface Echo Detection.

Area of Science:

• Ultrasound physics and signal processing
• Artificial intelligence in imaging
• Non-Destructive Testing (NDT) and medical diagnostics

Background:

• Ultrasound array imaging relies on accurate Time of Flight (TOF) measurements for focal law derivation.
• Surface echoes are critical for TOF determination but conventional detection methods are noise-sensitive.
• Robust techniques are needed to overcome limitations in traditional threshold crossing and peak search algorithms.

Purpose of the Study:

• To develop and evaluate a deep 3D Convolutional Neural Network (CNN) for precise surface echo detection in ultrasound Full Matrix Capture (FMC) data.
• To assess the performance of the proposed CNN model against established TOF estimation methods.
• To demonstrate the potential of AI in enhancing the accuracy of ultrasound imaging.

Main Methods:

• A deep 3D CNN, named DeepEcho3D, was designed for surface echo detection.
• The CNN was trained using FMC ultrasound signals from a matrix array and reference components.
• Precise probe positioning was ensured by a robotic arm setup, and labeled data was generated using theoretical TOFs.

Main Results:

• The DeepEcho3D model demonstrated high alignment with ground truth TOF values.
• The CNN significantly reduced TOF estimation outliers by up to 98% compared to conventional methods.
• The proposed method shows superior robustness in noisy ultrasound signal environments.

Conclusions:

• DeepEcho3D offers a highly accurate and robust solution for surface echo detection in ultrasound imaging.
• The AI-driven approach significantly improves TOF estimation, benefiting NDT and medical imaging applications.
• This study highlights the efficacy of deep learning in advancing ultrasound data analysis and interpretation.