The effect of preprocessing filters on predictive performance in radiomics.

Area of Science:

• Radiomics
• Machine Learning
• Personalized Medicine

Background:

• Radiomics, a noninvasive machine learning method, supports personalized medicine.
• Common preprocessing filters (e.g., wavelet, Laplacian-of-Gaussian) are thought to improve predictive performance.
• However, filters increase feature complexity and potential correlations, challenging machine learning models.

Purpose of the Study:

• To investigate the impact of preprocessing filters on radiomic predictive performance.
• To evaluate whether common preprocessing filters enhance or hinder machine learning model accuracy in radiomics.

Main Methods:

• Utilized seven public radiomic datasets.
• Compared models with and without features preprocessed by eight different filters.
• Employed five feature selection methods and five classifiers.
• Measured performance using Area Under the Curve (AUC-ROC) with nested, stratified 10-fold cross-validation.

Main Results:

• Applying all preprocessing filters significantly improved AUC-ROC by up to 0.08 (p = 0.024) compared to original features.
• Some datasets showed non-significant decreases in AUC-ROC (-0.04 to -0.10).
• Filter tuning further improved results by up to 0.1, with one dataset showing a significant improvement (p = 0.023).

Conclusions:

• Image preprocessing filters significantly impact radiomic predictive performance.
• The use of preprocessing filters is recommended in radiomic studies to potentially enhance predictive accuracy.