Functional Principal Component Analysis and Randomized Sparse Clustering Algorithm for Medical Image Analysis.

Area of Science:

• Medical image analysis
• Computational biology
• Biostatistics

Background:

• Advancements in sensor technology generate large medical image datasets, crucial for disease diagnosis and outcome prediction.
• High dimensionality of medical images presents significant computational and methodological challenges for feature representation and selection in cluster analysis.
• Current sparse clustering methods using lasso-type penalties for feature selection are sensitive to parameter and threshold choices, impacting accuracy.

Purpose of the Study:

• To address the challenges in medical image cluster analysis posed by high-dimensional data.
• To develop an accurate and efficient feature selection method for image clustering.
• To improve the characterization of disease progression and clinical outcomes through enhanced image analysis.

Main Methods:

• Extended functional principal component analysis (FPCA) from one to two dimensions to capture spatial variations in image signals.
• Developed a novel randomized algorithm for accurate feature selection in image clustering.
• Applied the proposed method to liver and kidney cancer histology image data from The Cancer Genome Atlas (TCGA) database.

Main Results:

• The proposed randomized feature selection method, when combined with 2D FPCA, effectively handles redundant features in medical images.
• Demonstrated substantial performance improvements compared to current sparse clustering algorithms in image cluster analysis.
• Validated the efficacy of the method on real-world histology data from liver and kidney cancer studies.

Conclusions:

• The randomized feature selection coupled with functional principal component analysis offers a robust solution for high-dimensional medical image clustering.
• This approach overcomes limitations of traditional sparse clustering methods, particularly in determining penalty parameters and thresholds.
• The findings suggest a significant advancement in the computational analysis of medical images for improved diagnostic and prognostic capabilities.