MR image segmentation and bias field estimation based on coherent local intensity clustering with total variation...

Area of Science:

• Medical Imaging
• Image Processing
• Computational Neuroscience

Background:

• Existing magnetic resonance (MR) image segmentation algorithms often address tissue segmentation, bias correction, and noise reduction separately.
• A unified approach is needed to simultaneously improve accuracy and efficiency in MR image analysis.

Purpose of the Study:

• To develop a novel unified framework for MR image analysis.
• To integrate brain tissue segmentation, bias field correction, and noise reduction into a single energy model.

Main Methods:

• A new energy model incorporating a total variation term into the coherent local intensity clustering criterion.
• Introduction of auxiliary variables to address the non-convexity of the membership functions.
• Application of Chambolle's fast dual projection method for simultaneous optimal segmentation and bias field estimation via reciprocal iteration.

Main Results:

• The proposed unified method significantly outperforms three baseline methods in both tissue segmentation and bias correction.
• Effective noise reduction is achieved, particularly on highly noise-corrupted MR images.
• The algorithm exhibits fast convergence, leading to reduced computation time and robustness to parameter settings.

Conclusions:

• The unified framework offers a salient advantage over existing methods for MR image analysis.
• Simultaneous integration of segmentation, bias correction, and noise reduction enhances overall image quality and analytical accuracy.
• The method's efficiency and robustness make it a valuable tool for processing challenging MR datasets.