Deep learning-based motion tracking using ultrasound images.

Area of Science:

• Medical Imaging
• Radiation Oncology
• Artificial Intelligence

Background:

• Ultrasound (US) imaging provides radiation-free, video-rate volumetric imaging.
• Intra-fraction motion tracking is crucial for effective radiation therapy.
• Challenges exist in precise motion tracking using US imaging.

Purpose of the Study:

• To develop a deep learning-based method for accurate motion tracking in ultrasound (US) imaging.
• To address the challenges of real-time motion tracking for radiation therapy applications.

Main Methods:

• A Markov-like network, implemented with generative adversarial networks, was developed.
• The network extracts features from sequential US frames to estimate deformation vector fields (DVFs).
• Landmark positions are determined by registering tracked and untracked frames using estimated DVFs.

Main Results:

• The method was evaluated on the MICCAI CLUST 2015 and CAMUS datasets.
• Mean tracking errors of 0.70 ± 0.38 mm (2D) and 1.71 ± 0.84 mm (3D) were achieved on the CLUST dataset.
• A mean tracking error of 0.54 ± 1.24 mm was achieved for left atrium landmarks on the CAMUS dataset.

Conclusions:

• A novel deep learning-based ultrasound motion tracking algorithm was demonstrated.
• The method offers real-time, millimeter-level tumor motion prediction.
• This technique holds potential for routine tumor motion management in radiation therapy.