The effect of feature normalization methods in radiomics.

Area of Science:

• Radiomics
• Medical Image Analysis
• Machine Learning in Healthcare

Background:

• Radiomics utilizes quantitative features extracted from medical images.
• Various feature normalization techniques are applied in radiomics, but their impact is not fully understood.
• Understanding normalization's effect is crucial for reliable radiomic model development.

Purpose of the Study:

• To evaluate the impact of different feature normalization methods on radiomic model performance.
• To assess how normalization affects feature selection and model calibration.
• To determine if normalization before cross-validation introduces bias.

Main Methods:

• Compared seven normalization methods across fifteen public radiomics datasets.
• Utilized four feature selection and classifier algorithms with cross-validation.
• Measured Area Under the Curve (AUC), feature selection agreement, and model calibration.
• Assessed bias introduced by pre-cross-validation normalization.

Main Results:

• Normalization methods showed minor average differences in AUC, with z-score performing best.
• Significant performance variations were observed across different datasets.
• Feature selection agreement between methods was low (≤62%).
• Normalization before cross-validation did not introduce significant bias.

Conclusions:

• Feature normalization significantly influences radiomic model performance and feature selection, with dataset-specific effects.
• The choice of normalization method complicates feature interpretation.
• While z-score generally performed well, dataset variability necessitates careful method selection.