Radiomics analysis using stability selection supervised component analysis for right-censored survival data.

Area of Science:

• Radiomics and Machine Learning
• Biomedical Imaging Analysis
• Cancer Prognostics

Background:

• Radiomics extracts quantitative features from medical images for prognosis and treatment prediction.
• Current radiomics studies often use generic algorithms not optimized for radiomics data.
• There's a need for specialized methods to identify robust prognostic markers from complex radiomics datasets.

Purpose of the Study:

• To develop a novel machine learning approach for radiomics data analysis.
• To identify stable and significant imaging features associated with survival outcomes.
• To improve the prediction of patient survival using noninvasive imaging biomarkers.

Main Methods:

• Proposed stability selection supervised principal component analysis (SSSuperPCA) for feature selection and dimension reduction.
• Applied the method to radiomics data for right-censored survival outcomes.
• Controlled the per-family error rate for robust feature identification.

Main Results:

• SSSuperPCA successfully identified stable features highly correlated with survival outcomes.
• The method demonstrated a competitive edge over existing algorithms in real-world cancer datasets (non-small cell lung cancer, head and neck cancer).
• Validated performance through simulations and clinical data analysis.

Conclusions:

• SSSuperPCA offers a powerful and meaningful approach for prognostic marker discovery in radiomics.
• This novel method enhances the predictive accuracy of survival outcomes from biomedical images.
• It provides a valuable tool for personalized cancer therapy and noninvasive patient management.