Semi-supervised cluster analysis of imaging data.

Area of Science:

• Medical Imaging Analysis
• Computational Biology
• Machine Learning

Background:

• Heterogeneous image datasets often contain distinct subpopulations that are difficult to identify.
• Limited supervision can guide clustering algorithms for more relevant discoveries.
• Analyzing populations of healthy older adults requires robust methods for subgroup identification.

Purpose of the Study:

• To develop a semi-supervised clustering framework for discovering coherent subpopulations in heterogeneous image sets.
• To create a method for detecting local regional differences in labeled image sets.
• To assess the framework's utility in analyzing brain MR images of older adults and identifying deviations from a normal state.

Main Methods:

• A segmentation-based method identifies locations of regional differences in labeled image sets.
• A Principal Component Analysis (PCA) model of local image appearance is estimated and ranked for clustering relevance.
• An incremental k-means-like algorithm discovers novel categories within test image sets.

Main Results:

• The framework successfully identifies meaningful categories in test image sets.
• Validation on synthetic and real brain image datasets demonstrates effectiveness.
• A cluster-based measure of pathology was developed, reflecting deviations from a cognitively stable state.
• Clustering results showed good correlation with clinical data.

Conclusions:

• The proposed semi-supervised clustering framework effectively discovers subpopulations in heterogeneous image data.
• The method provides a valuable tool for analyzing brain MR images and assessing deviations from normal states.
• The approach shows promise for correlating imaging findings with clinical data in aging populations.