The Challenge of Choosing the Best Classification Method in Radiomic Analyses: Recommendations and Applications to...

Area of Science:

• Radiomics and Machine Learning in Medical Imaging

Background:

• Radiomics leverages high-dimensional imaging features for tumor characterization and outcome prediction.
• Algorithm selection critically influences radiomic analysis and predictive accuracy.
• Identifying optimal machine learning methods is essential for reliable radiomic applications.

Purpose of the Study:

• To identify suitable machine learning approaches for radiomic-based binary predictions.
• To evaluate the impact of sample size, outcome balancing, and feature-outcome association strength on prediction performance.
• To assess the effectiveness of various feature selection (FS) methods combined with classifiers.

Main Methods:

• Simulated data based on 168 radiomic features from CT scans of 270 Non-Small-Cell Lung Cancer (NSCLC) patients were used.
• Six classifiers were evaluated in combination with six FS methods.
• Performance metrics included Area Under the Receiver Operating Characteristics Curves (AUC), sensitivity, and specificity.

Main Results:

• Tree-based classifiers (Random Forest, Extreme Gradient Boosting) demonstrated strong performance (AUC ≥ 0.73) across various FS methods and association strengths.
• These classifiers offered the best balance between sensitivity and specificity.
• Performance was generally lower and more variable with small sample sizes.
• FS methods did not consistently improve classifier performance.

Conclusions:

• Random Forest and Extreme Gradient Boosting are recommended for radiomic binary predictions, particularly for NSCLC lymph node status.
• Careful consideration of sample size, outcome balancing, and feature-outcome association is crucial when selecting FS and classifiers in radiomic studies.