Temporal contexts for motion tracking in ultrasound sequences with information bottleneck

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep convolutional neural networks (CNNs) are common for ultrasound tracking but often overlook temporal context.
• Ignoring temporal information hinders accurate perception of target motion in ultrasound sequences.

Purpose of the Study:

• To develop a sophisticated method for ultrasound sequence tracking that fully leverages temporal contexts.
• Integrate information bottleneck to refine feature extraction and similarity graph construction.

Main Methods:

• Proposed a tracker combining an online temporal adaptive convolutional neural network (TAdaCNN) for feature extraction.
• Incorporated information bottleneck (IB) to discard irrelevant information and improve tracking accuracy.
• Utilized a temporal adaptive transformer (TA-Trans) for encoding temporal knowledge and refining similarity graphs.
• Trained and evaluated on the CLUST 2015 dataset, comparing with 13 state-of-the-art methods.

Main Results:

• Achieved a mean tracking error (TE) of 0.81 ± 0.74 mm on the CLUST 2015 dataset.
• Demonstrated tracking speeds of 41-63 frames per second (fps).
• Outperformed 13 state-of-the-art methods in accuracy and robustness.

Conclusions:

• The proposed integrated workflow offers a new approach to ultrasound sequence motion tracking.
• The model exhibits excellent accuracy and robustness for real-time motion estimation.
• Provides reliable motion estimation crucial for ultrasound-guided radiation therapy.