Physics informed contour selection for rapid image segmentation.

Area of Science:

• Medical Imaging
• Computational Science
• Artificial Intelligence

Background:

• Deep learning models for image segmentation require extensive, high-quality annotations, posing a significant training challenge.
• Existing active contour models (snakes) often rely on complex mathematical derivations and edge-based loss functions.

Purpose of the Study:

• Introduce Physics Informed Contour Selection (PICS), a novel algorithm for rapid, interpretable image segmentation.
• Develop a method that reduces reliance on labeled data and enhances computational efficiency.

Main Methods:

• PICS utilizes cubic splines for computational lightness and employs physics-informed neural networks (PINNs) principles.
• The algorithm directly minimizes a region-based loss functional, bypassing traditional Euler-Lagrange equations.
• A novel convexity-preserving loss term is proposed for encoding prior shape information in 3D segmentation.

Main Results:

• PICS demonstrates fast and computationally lightweight 3D segmentation without requiring labeled data.
• The algorithm achieves effective 3D segmentation of the left ventricle in cardiac datasets.
• PICS successfully encodes prior shape information through a new loss term.

Conclusions:

• PICS offers a novel approach to image segmentation, combining physics-based principles with efficient computational methods.
• The algorithm presents advancements in network architecture, transfer learning, and physics-inspired losses.
• PICS shows significant potential for improving 3D image segmentation tasks, particularly in medical applications.