3d freehand ultrasound reconstruction by reference-based point cloud registration.

Area of Science:

• Medical Imaging
• Computer Vision
• Geometric Deep Learning

Background:

• Estimating trajectories from freehand ultrasound is challenging.
• Current point cloud registration methods struggle with realistic examination conditions like sparse labels and partial overlap.

Purpose of the Study:

• To develop an improved pipeline for trajectory estimation from freehand ultrasound data.
• To enhance point cloud registration by incorporating heatmap predictions and multi-stage refinement.

Main Methods:

• A three-stage pipeline: Free Point Transformer for pre-registration, HeatReg with support points for refinement, and instance optimization for displacement enhancement.
• Utilized prior knowledge and gradient keypoints to expand point sets.
• Evaluated on 42 forearm ultrasound sweeps with optical ground-truth tracking.

Main Results:

• The proposed pipeline effectively registers freehand intra-patient ultrasound sweeps.
• The combined approach (Free Point Transformer + HeatReg) improved upon the baseline by 40% (0.96 mm mean directed surface distance).
• Further refinement enhanced registration accuracy and trajectory estimation.

Conclusions:

• The developed techniques enable and improve point cloud registration for freehand ultrasound reconstruction.
• Heatmap incorporation and multi-stage predictions offer significant theoretical and practical advantages.