Reproducibility in Radiomics: A Comparison of Feature Extraction Methods and Two Independent Datasets.

Area of Science:

• Medical Imaging Analysis
• Radiomics Feature Reproducibility
• Computational Pathology

Background:

• Radiomics extracts quantitative imaging features for clinical applications.
• Reproducibility of radiomics features across studies is a significant concern.
• Standardized validation of feature extraction methods is essential for clinical translation.

Purpose of the Study:

• To assess the reproducibility of CT texture features between MATLAB and Pyradiomics implementations.
• To identify radiomics features that are reproducible across different datasets and software.
• To validate a methodology for evaluating radiomics feature reproducibility.

Main Methods:

• Comparison of 43 common radiomics features from MATLAB and Pyradiomics.
• Utilized independent CT datasets (RIDER and HN1) with Gross Tumor Volume (GTV).
• Calculated Spearman's rank correlation coefficient (rs) to assess feature rank ordering consistency.

Main Results:

• 29 out of 43 features demonstrated high reproducibility (rs > 0.8) between MATLAB and Pyradiomics.
• 18 features were common across lung and head and neck cancer datasets, suggesting potential site-agnostic utility.
• The study validated a robust methodology for assessing radiomics reproducibility.

Conclusions:

• A subset of radiomics features exhibits high reproducibility between common software tools.
• Reproducibility is a critical criterion for selecting reliable radiomics features for clinical use.
• The findings support the use of validated, reproducible radiomics features for treatment stratification and outcome prediction.