Machine Learning methods for Quantitative Radiomic Biomarkers.

Area of Science:

• Medical Imaging
• Radiomics
• Machine Learning
• Oncology

Background:

• Radiomics quantifies tumor characteristics from medical images.
• Accurate machine learning (ML) is vital for clinical radiomics.
• Predicting patient outcomes non-invasively is a key goal in cancer care.

Purpose of the Study:

• To evaluate the performance and stability of various feature selection and classification methods for predicting overall survival in lung cancer patients using radiomics.
• To identify optimal ML approaches for robust radiomic biomarker development.

Main Methods:

• Extracted 440 radiomic features from pre-treatment CT scans of 464 lung cancer patients.
• Evaluated 14 feature selection methods and 12 classification methods using independent training and validation cohorts.
• Utilized publicly available ML implementations with reported parameters for unbiased evaluation.

Main Results:

• The Wilcoxon test (WLCX) feature selection and random forest (RF) classification demonstrated the highest prognostic performance and stability.
• WLCX achieved an AUC of 0.65 ± 0.02 with 0.84 ± 0.05 stability.
• RF achieved an AUC of 0.66 ± 0.03 with 3.52% RSD, and classification method choice significantly impacted performance variation.

Conclusions:

• Optimal machine learning methods are crucial for developing stable and clinically relevant radiomic biomarkers.
• This research provides insights into selecting robust ML approaches for radiomics in lung cancer.
• Radiomics offers a non-invasive method for quantifying and monitoring tumor phenotypes in clinical practice.