Robust vs. Non-robust radiomic features: the quest for optimal machine learning models using phantom and clinical...

Area of Science:

• Radiomics
• Medical Imaging Analysis
• Machine Learning in Oncology

Background:

• Radiomic features are sensitive to motion artifacts, impacting the reproducibility and reliability of predictions.
• Lymphovascular invasion (LVI) is a critical prognostic factor in non-small cell lung cancer (NSCLC).
• Developing robust methods for LVI prediction is essential for improving patient outcomes.

Purpose of the Study:

• To select radiomic features robust against lung motion using a phantom study.
• To evaluate the performance of robust features in predicting LVI in NSCLC using machine learning.
• To compare the efficacy of robust features against conventional methods.

Main Methods:

• A lung phantom with simulated motion was used to identify radiomic features resistant to motion artifacts.
• 105 radiomic features were extracted from phantom and clinical datasets (n=126).
• Feature selection algorithms and machine learning classifiers were applied to predict LVI, comparing robust versus conventional features.

Main Results:

• Using robust features significantly increased prediction sensitivity, with a minor impact on accuracy and AUC in 12 out of 15 outcomes.
• The highest performance without robust features was 95% AUC, 67% accuracy, and 100% sensitivity (NB classifier, RFE FS).
• The highest performance with robust features was 92% AUC, 86% accuracy, and 100% sensitivity (NB classifier, Boruta FS).

Conclusions:

• Feature robustness against influential factors like motion is crucial for reliable radiomic studies.
• Selecting motion-robust features is a viable strategy to enhance the reproducibility of radiomic analyses.
• While slightly impacting accuracy and AUC, robust features substantially improve LVI prediction sensitivity.