Inconsistent CT NSCLC radiomics associated with feature selection methods, predictive models and related factors

Area of Science:

• Radiomics and Medical Imaging
• Oncology and Cancer Research
• Computational Biology and Bioinformatics

Background:

• Computed tomography (CT)-based radiomics shows promise for non-small cell lung cancer (NSCLC) prognostication.
• However, uncertainties related to feature selection and predictive modeling may affect the reliability of radiomic studies.

Purpose of the Study:

• To investigate the impact of feature selection methods, predictive models, and cohort characteristics on CT-based NSCLC radiomics.
• To identify factors contributing to variability and potential inconsistencies in radiomic analysis.

Main Methods:

• Retrospective analysis of CT images from 496 pre-treatment NSCLC patients.
• Radiomic features extracted using IBEX; five feature selection methods and seven predictive models were evaluated.
• Impact of cohort size (25%, 50%, 75%) and composition on feature selection and model performance (AUC) assessed.

Main Results:

• Feature selection results were inconsistent and dependent on cohort size and composition.
• Only 16 out of 2100 tested combinations achieved an Area Under the Curve (AUC) > 0.65.
• Significant variability observed, indicating no clear standardized approach for reliable CT NSCLC radiomics.

Conclusions:

• The choice of feature selection methods and predictive models significantly influences NSCLC radiomic study outcomes.
• Inconsistencies underscore the need for further investigation into standardization to enhance the reliability of radiomic biomarkers.
• Addressing these uncertainties is crucial for the clinical translation of CT-based radiomics in NSCLC.