Universal multi-factor feature selection method for radiomics-based brain tumor classification.

Area of Science:

• Medical imaging analysis
• Machine learning in oncology

Background:

• Brain tumor classification is crucial for treatment and prognosis.
• Radiomics analysis of medical images offers valuable insights.
• Current feature selection methods lack universality, hindering progress.

Purpose of the Study:

• To address the limitations of existing feature selection methods in radiomics for brain tumor classification.
• To propose a universal feature selection method, the triple-factor cascaded selection (TFCS).

Main Methods:

• TFCS utilizes three factors: feature-label correlation, feature interdependence, and feature role in the model.
• The method employs mutual information for initial feature selection and recursive feature elimination for refinement.
• Validation was performed on seven datasets across 13 brain tumor classification tasks.

Main Results:

• TFCS demonstrated excellent performance across all tasks, outperforming 13 related methods.
• The method showed superior classification performance, adaptability, and stability.
• TFCS achieved these results with reduced computation time and moderate parsimony.

Conclusions:

• The proposed TFCS method offers a universal and effective approach to feature selection in radiomics.
• Utilizing multiple factors in feature selection enhances performance and provides a foundation for future method development.
• This advancement can improve the accuracy and reliability of brain tumor classification, ultimately benefiting patient outcomes.