A Multi-Source Circular Geodesic Voting Model for Image Segmentation

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) excel in feature learning but lack geometric priors and global context.
• Variational methods offer geometric priors but require manual initialization and are sensitive to hyperparameters.
• Existing methods struggle with accuracy in complex imaging scenarios.

Purpose of the Study:

• To develop a novel image segmentation approach, PolarVoting, that integrates CNNs with geometric priors.
• To improve segmentation accuracy and robustness, especially in challenging imaging conditions.
• To leverage minimal path encoding and a modified circular geometric voting scheme.

Main Methods:

• Utilizing the PolarMask model to extract multiple source points for initialization.
• Constructing a voting score map using a modified circular geometric voting (CGV) scheme.
• Embedding global geometric information within the voting map for accurate segmentation.

Main Results:

• The PolarVoting method demonstrated superior performance compared to PolarMask and traditional single-source CGV models.
• Achieved enhanced segmentation accuracy and robustness across various datasets.
• Successfully delineated object boundaries in images with intensity inhomogeneity, noise, and complex backgrounds.

Conclusions:

• PolarVoting effectively integrates neural network representations with geometric priors for advanced image segmentation.
• The proposed method offers a significant advancement in addressing limitations of current segmentation techniques.
• PolarVoting shows promise for applications in computer vision and medical imaging requiring precise object delineation.