Label inference with registration and patch priors.

Area of Science:

• Medical image analysis
• Computer vision

Background:

• Accurate segmentation of anatomical structures is crucial in medical imaging.
• Labeling errors, especially around structural boundaries, pose a significant challenge.
• Existing non-rigid registration methods provide initial segmentation but often contain inaccuracies.

Purpose of the Study:

• To develop a novel label inference method to correct segmentation errors near structural boundaries.
• To improve the accuracy and reliability of medical image segmentation.
• To provide a computationally efficient solution for refining segmentation maps.

Main Methods:

• A novel label inference method integrating registration and patch priors.
• Utilizing signed distance functions derived from non-rigid registration for confidence evaluation.
• Employing a seed-based approach where confident pixels act as seeds for refining less confident candidate pixels.
• Encoding affinities including image lattice connections, signed distance-based registration prior, and patch priors from a warped atlas.

Main Results:

• The proposed method effectively remedies labeling errors around structural boundaries.
• Significant improvement in segmentation quality was observed.
• The method demonstrated high computational efficiency, requiring only seconds for refinement.
• Validation was performed on two publicly available datasets.

Conclusions:

• The novel label inference method offers a robust solution for correcting segmentation inaccuracies.
• Integration of registration and patch priors enhances segmentation accuracy.
• The method is efficient and significantly improves segmentation quality in medical imaging applications.