A Deep Learning-Based Automatic First-Arrival Picking Method for Ultrasound Sound-Speed Tomography.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Machine Learning

Background:

• Ultrasound sound-speed tomography (USST) offers non-ionizing, cost-effective 3D breast imaging for cancer diagnosis.
• Accurate first-arrival picking of transmission waves is critical for USST reconstruction quality.
• Existing picking methods lack accuracy and robustness against noise.

Purpose of the Study:

• To develop an improved first-arrival picking method for USST.
• To enhance the accuracy and robustness of wave picking using deep learning.
• To improve the overall quality of USST breast imaging.

Main Methods:

• Introduction of a self-attention mechanism into a bidirectional long short-term memory (BLSTM) network, creating the SAT-BLSTM network.
• The SAT-BLSTM network predicts the probability of first-arrival times, selecting the time with maximum probability.
• Validation through numerical simulations and a prototype experiment.

Main Results:

• SAT-BLSTM achieved the lowest mean absolute errors (MAEs) in numerical simulations across various signal-to-noise ratios (SNRs): 48 ns (50 dB), 49 ns (30 dB), and 76 ns (15 dB).
• In prototype experiments, SAT-BLSTM demonstrated superior performance with an MAE of 94 ns compared to BLSTM (111 ns) and AIC (410 ns).
• The proposed method significantly outperformed traditional Akaike information criterion (AIC) and standard BLSTM methods.

Conclusions:

• The SAT-BLSTM network offers a significant advancement in first-arrival picking for USST.
• This improved picking accuracy leads to enhanced USST image reconstruction and potentially more reliable breast cancer diagnosis.
• The self-attention mechanism integration boosts robustness and accuracy in noisy environments.