A novel rotationally invariant region-based hidden Markov model for efficient 3-D image segmentation

Area of Science:

• Medical imaging
• Computer vision
• Machine learning

Background:

• Unsupervised 3-D image segmentation is crucial for medical data analysis.
• Current Hidden Markov Model (HMM) approaches often use inefficient data representations (voxels, grids).
• Segmentation results can be sensitive to object orientation, impacting variability.

Purpose of the Study:

• To introduce a novel 3-D region-based hidden Markov model (rbHMM) for efficient unsupervised 3-D image segmentation.
• To develop a tree-structured parameter estimation algorithm for improved segmentation accuracy and rotation invariance.
• To demonstrate the rbHMM's superiority over existing grid-based HMMs.

Main Methods:

• Developed a region-based HMM (rbHMM) with an efficient image data representation.
• Implemented a novel tree-structured parameter estimation algorithm for locally optimal, rotation-invariant labeling.
• Validated the rbHMM on synthetic geometric shapes and simulated/clinical brain MRI scans.

Main Results:

• rbHMM achieved faster optimization compared to grid-based HMMs.
• Segmentation results on geometric shapes were accurate and rotation-invariant.
• Brain MRI segmentation showed improved robustness and accuracy, with enhanced Dice similarity indices for white and gray matter (4.60% and 7.71%, respectively).

Conclusions:

• The proposed rbHMM offers an efficient and accurate solution for unsupervised 3-D image segmentation.
• The rotation-invariant property is particularly valuable for medical imaging, reducing methodological variability.
• rbHMM demonstrates significant advantages over traditional grid-based HMMs in both synthetic and real-world medical imaging data.