Benchmarking Feature Selection Methods in Radiomics.

Area of Science:

• Radiomics
• Medical Imaging Analysis
• Machine Learning in Healthcare

Background:

• High-dimensional data in radiomics arises from small sample sizes and numerous extracted features.
• Feature selection is crucial for removing redundant and irrelevant features in radiomic datasets.
• Understanding the performance of various feature selection algorithms is key for effective radiomic studies.

Purpose of the Study:

• To evaluate the performance of 29 feature selection algorithms in radiomic studies.
• To compare algorithms based on training times, feature stability, and ranking similarity.
• To assess the predictive performance of feature selection methods using the area under the receiver operating characteristic curve.

Main Methods:

• Evaluated 29 feature selection algorithms and 10 classifiers on 10 public radiomic datasets.
• Compared methods for training time, feature selection stability, and pairwise similarity (ranking).
• Measured predictive performance using the area under the receiver operating characteristic curve (AUC) of the best classifier.

Main Results:

• Significant differences observed in training times, stability, and similarity across feature selection methods.
• No single feature selection method consistently outperformed others in predictive performance.
• Simpler feature selection methods demonstrated greater stability without compromising predictive accuracy (AUC).

Conclusions:

• Simpler feature selection methods are more stable and perform comparably to complex methods in radiomics.
• Analysis of variance (ANOVA), least absolute shrinkage and selection operator (LASSO), and minimum redundancy, maximum relevance (mRMR) ensemble are recommended.
• These recommended methods showed superior predictive performance compared to most other evaluated feature selection techniques.