Benchmarking feature projection methods in radiomics.

Area of Science:

• Medical Imaging Analysis
• Radiomics
• Machine Learning in Healthcare

Background:

• Radiomics utilizes quantitative features from medical images for clinical outcome prediction.
• Feature selection is standard, aiming to reduce dimensionality and enhance model interpretability.
• Feature projection methods are less common due to interpretability concerns, despite potential performance benefits.

Purpose of the Study:

• To compare the predictive performance of feature projection methods versus feature selection in radiomics.
• To evaluate if projection methods can outperform traditional selection techniques.
• To assess the impact on key performance metrics like AUC, AUPRC, and F-scores.

Main Methods:

• Trained models on 50 diverse radiomic datasets (CT/MRI) for binary classification tasks.
• Compared nine feature projection methods (e.g., PCA, NMF) against nine feature selection methods (e.g., MRMRe, ET, LASSO).
• Utilized nested, stratified 5-fold cross-validation with 10 repeats for robust evaluation.

Main Results:

• Feature selection methods, particularly ET, MRMRe, Boruta, and LASSO, generally yielded the highest overall performance.
• Performance varied significantly across datasets; NMF occasionally outperformed all selection methods.
• The average performance difference between selection and projection methods was negligible and not statistically significant.

Conclusions:

• Feature selection methods remain the primary choice for typical radiomics studies.
• Feature projection methods warrant consideration for potentially maximizing predictive performance.
• Methodological choice should balance interpretability with the pursuit of optimal predictive accuracy.