Finding Reproducible and Prognostic Radiomic Features in Variable Slice Thickness Contrast Enhanced CT of Colorectal...

Area of Science:

• Radiology
• Oncology
• Medical Imaging

Background:

• Reproducibility and prognostic value of radiomic signatures are essential for clinical adoption in personalized medicine.
• Radiomics involves extracting quantitative features from medical images to correlate with clinical outcomes.
• Colorectal liver metastases (CRLM) present a significant challenge where quantitative imaging biomarkers can aid treatment decisions.

Purpose of the Study:

• To assess the reproducibility and prognostic significance of radiomic features from contrast-enhanced CT scans in patients with CRLM.
• To evaluate radiomic features from both liver parenchyma and the largest liver metastases.
• To determine optimal feature extraction settings for reproducible and prognostically valuable radiomic signatures.

Main Methods:

• Analyzed prospective (n=81) and public (n=197) cohorts of CRLM patients.
• Extracted 93 standard radiomic features using eight different settings from CT images with varying slice thicknesses.
• Assessed feature reproducibility using Concordance Correlation Coefficient (CCC) and prognostic value for overall survival.

Main Results:

• Optimal feature extraction settings varied significantly based on region of interest and feature type.
• A predictive model using features selected for reproducibility (CCC ≥ 0.85) achieved performance equivalent to a model using all features (C-index ≈ 0.63).
• Pooling features from all extraction settings and then thresholding for reproducibility before selection proved effective.

Conclusions:

• A data-driven approach to radiomic feature extraction and selection is recommended.
• Prioritizing feature reproducibility enhances the development of reliable prognostic models for CRLM.
• Inclusion of numerous features followed by reproducibility-based filtering is a robust strategy for feature selection.