Deep learning-based high precision 3D ultrasound imaging for large size organ

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Free-hand 3D ultrasound imaging offers advantages in device simplicity and user-friendliness.
• Deep learning networks are used for spatial coordinate prediction in this imaging modality.
• A key limitation is cumulative errors in spatial transformation prediction, leading to 3D image distortion, especially for large organs.

Purpose of the Study:

• To develop a novel labeling strategy to enhance deep learning network prediction accuracy in free-hand 3D ultrasound.
• To reduce cumulative errors in 3D ultrasound reconstruction.
• To improve the clinical applicability of 3D ultrasound imaging.

Main Methods:

• Proposed a labeling strategy based on the ultrasound image coordinate system.
• Implemented pre-planning of scanning trajectories to guide network prediction.
• Applied the method to spinal 3D ultrasound imaging in healthy volunteers and scoliosis patients.

Main Results:

• The proposed labeling strategy enhanced network prediction accuracy.
• Pre-planning trajectories significantly reduced cumulative error.
• Prediction accuracy improved by approximately 40%, and cumulative error was reduced by nearly 80% compared to existing methods.

Conclusions:

• The novel labeling strategy and trajectory guidance effectively reduce cumulative errors in free-hand 3D ultrasound.
• This approach shows significant potential for improving 3D reconstruction accuracy and facilitating wider clinical adoption of 3D ultrasound.
• The method is adaptable to various deep learning networks and tissue types.