One Shot Segmentation: Unifying Rigid Detection and Non-Rigid Segmentation Using Elastic Regularization.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Machine Learning

Background:

• Accurate segmentation of deformable objects in image sequences is challenging.
• Existing methods often use sequential rigid detection and non-rigid segmentation steps.
• Deformable objects undergo both rigid and non-rigid transformations, complicating segmentation.

Purpose of the Study:

• To propose a novel, unified one-shot approach for non-rigid segmentation of deformable objects.
• To integrate rigid detection and non-rigid segmentation into a single framework.
• To reduce computational complexity and training data requirements.

Main Methods:

• A sparse, low-dimensional manifold represents object deformations.
• The manifold partitions training data into patches for segmentation proposals.
• Deep Belief Networks (DBN) ensemble classifiers merge proposals and estimate final segmentation.
• Elastic regularization is employed for unified segmentation.

Main Results:

• The proposed method achieves accurate segmentation for left ventricle endocardial and lip segmentation.
• Demonstrates reduced search space dimensionality compared to sequential methods.
• Requires smaller annotated training sets for DBN model estimation.
• Maintains segmentation accuracy while reducing complexity.

Conclusions:

• The unified one-shot framework effectively segments deformable objects with both rigid and non-rigid transformations.
• Ensemble classification enhances segmentation robustness.
• The approach offers significant reductions in search space and training complexity.
• Validated on ultrasound and facial image datasets.