Unsupervised Brain Lesion Segmentation Using Posterior Distributions Learned by Subspace-Based Generative Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Unsupervised brain lesion segmentation requires learning normative distributions from healthy subjects to reduce reliance on labeled data.
• High dimensionality poses a challenge when modeling spatial dependencies in image pixels as correlated random variables.

Purpose of the Study:

• To propose a subspace-based deep generative model for learning posterior normal distributions in brain images.
• To enhance unsupervised brain lesion segmentation by effectively capturing spatial-intensity and spatial-structure variations.

Main Methods:

• Utilized probabilistic subspace models to capture spatial-intensity and spatial-structure distributions from healthy brain images.
• Employed subspace coefficients as random variables with learned eigen-images and eigen-density functions.
• Integrated subspace-based generative models and Bayesian analysis for posterior distribution estimation.
• Applied an unsupervised fusion classifier to combine posterior and likelihood features for segmentation.

Main Results:

• The proposed model effectively captures prior spatial-intensity and spatial-structure variations.
• Demonstrated superior segmentation accuracy and robustness on simulated and real lesion data (tumor, multiple sclerosis, stroke).
• Outperformed existing state-of-the-art unsupervised segmentation methods.

Conclusions:

• The subspace-based deep generative model offers a promising approach for unsupervised brain lesion segmentation.
• The method exhibits enhanced generalization capabilities, reducing dependency on lesion-labeled datasets.
• Holds significant potential for improving clinical applications in brain lesion delineation.