Variability Regularized Feature Selection (VaRFS) for optimal identification of robust and discriminable features...

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Radiomics and Computational Pathology

Background:

• Machine learning models using computerized features from medical imaging show promise for disease prediction and prognosis.
• Model performance is often limited by feature robustness to variations in imaging institutions and acquisition protocols.

Purpose of the Study:

• To propose Variability Regularized Feature Selection (VaRFS), a novel framework to enhance feature selection for medical imaging.
• To identify imaging features that are discriminative between outcome groups and generalizable across different imaging conditions.

Main Methods:

• VaRFS integrates feature variability into a regularization term within the Least Absolute Shrinkage and Selection Operator (LASSO) framework.
• A novel sparse regularization strategy is employed, with confirmed convergence guarantees and an accelerated proximal variant for computational efficiency.
• The framework was evaluated on over 700 multi-institutional imaging datasets across five clinical applications.

Main Results:

• VaRFS consistently achieved higher classifier Area Under the Curve (AUC) values in hold-out validation compared to three conventional feature selection methods.
• The method demonstrated effectiveness in balancing feature reproducibility, sparsity, and discriminability.
• Successful application across disease detection, treatment response characterization, and risk stratification was shown.

Conclusions:

• Variability Regularized Feature Selection (VaRFS) offers a robust approach for selecting generalizable and discriminative features from medical imaging data.
• This framework improves the reliability and performance of machine learning models in clinical applications.
• VaRFS addresses the critical challenge of imaging variability, paving the way for more dependable AI-driven diagnostics and prognostics.