Simultaneous segmentation and multiresolution nonrigid atlas registration.

Area of Science:

• Medical image analysis
• Computational anatomy
• Computer vision

Background:

• Medical image segmentation and registration are crucial for clinical applications.
• Iterative approaches for combined tasks may not yield optimal results.
• Existing methods treat segmentation and registration separately, limiting performance.

Purpose of the Study:

• To develop a novel Markov random field (MRF)-based approach for simultaneous medical image segmentation and nonrigid atlas registration.
• To formulate the simultaneous registration and segmentation (SRS) problem as a maximum a-posteriori (MAP) problem.
• To improve accuracy and robustness compared to iterative methods.

Main Methods:

• Formulated simultaneous registration and segmentation (SRS) as a maximum a-posteriori (MAP) problem.
• Decomposed probabilities for MAP inference using MRFs.
• Employed an efficient hierarchical implementation for coarse-to-fine registration and pixel-level segmentation.

Main Results:

• Evaluated on 3D CT mandibular bone and 2D MRI corpus callosum segmentation.
• Demonstrated superior accuracy using Dice overlap and surface distance metrics compared to iterative approaches.
• Showed SRS is less sensitive to registration errors than iterative methods.

Conclusions:

• The proposed SRS method offers a more accurate and robust solution for combined medical image segmentation and registration.
• Simultaneous optimization outperforms iterative, separate optimization for these tasks.
• The MRF-based SRS approach is effective for diverse medical imaging datasets.