Shape-constrained Gaussian process regression for surface reconstruction and multimodal, non-rigid image...

Area of Science:

• Medical image analysis
• Computer vision
• Computational geometry

Background:

• Image registration and surface reconstruction are crucial for medical diagnosis and scientific visualization.
• Existing methods often struggle with accuracy, smoothness, and handling complex deformations.

Purpose of the Study:

• To develop a robust statistical framework for curve-based image registration and surface reconstruction.
• To accurately estimate dense deformation fields from sparse landmark curves.
• To ensure spatial smoothness of the estimated deformations.

Main Methods:

• A novel statistical framework utilizing elastic alignment of parameterized curves to compute local deformations.
• Employing a Gaussian random field model to estimate the full deformation vector field.
• Statistical estimation via maximum likelihood and Bayesian inference (Markov Chain Monte Carlo sampling).

Main Results:

• The proposed method achieves accurate matching of corresponding curve regions.
• Generated deformations are spatially smooth across the entire domain.
• Successful qualitative and quantitative evaluations on synthetic and real data.

Conclusions:

• The developed statistical framework provides an accurate and robust approach for image registration and surface reconstruction.
• The method is applicable to diverse tasks, including multimodal medical image registration and 3D surface reconstruction of anatomical and artifactual objects.