Self-supervised MRI tissue segmentation by discriminative clustering.

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Image Analysis

Background:

• Accurate identification of tissue transitions in brain imaging is crucial for studying brain lesions.
• Existing signal processing methods often require extensive prior information, limiting their versatility.
• There is a need for adaptable segmentation methods capable of handling diverse tissue types.

Purpose of the Study:

• To propose a novel, versatile tissue segmentation method for brain imaging analysis.
• To develop a self-supervised machine learning approach that avoids reliance on prior information.
• To enable precise characterization of brain lesion evolution through voxel-wise tissue probabilities.

Main Methods:

• A discriminative strategy within a self-supervised machine learning framework was employed.
• The method was designed to be versatile, applicable to various tissue types without prior assumptions.
• Tissue probabilities were generated for each voxel, facilitating detailed analysis.

Main Results:

• The proposed method demonstrated high accuracy in segmenting brain tissues.
• Validation using simulated and real benchmark data confirmed its effectiveness.
• Performance comparison showed favorable results against existing segmentation algorithms.

Conclusions:

• The novel self-supervised method offers a versatile and accurate approach to brain tissue segmentation.
• Its ability to work without prior information enhances its applicability across different imaging scenarios.
• The voxel-wise probability outputs are valuable for characterizing brain lesion dynamics.