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

TL;DR

This study evaluates the reproducibility and prognostic significance of radiomic features from contrast-enhanced CT scans of colorectal liver metastases, emphasizing the importance of reproducibility in feature selection for clinical prediction models.

Contribution

It demonstrates that reproducibility-based feature filtering can maintain predictive performance while improving model robustness in radiomic analysis.

Findings

Reproducibility of radiomic features varies by region and feature type.

Pooling features across extraction settings with reproducibility filtering yields comparable prognostic models.

Reproducibility-based feature selection enhances model robustness without sacrificing accuracy.

Abstract

Establishing the reproducibility of radiomic signatures is a critical step in the path to clinical adoption of quantitative imaging biomarkers; however, radiomic signatures must also be meaningfully related to an outcome of clinical importance to be of value for personalized medicine. In this study, we analyze both the reproducibility and prognostic value of radiomic features extracted from the liver parenchyma and largest liver metastases in contrast enhanced CT scans of patients with colorectal liver metastases (CRLM). A prospective cohort of 81 patients from two major US cancer centers was used to establish the reproducibility of radiomic features extracted from images reconstructed with different slice thicknesses. A publicly available, single-center cohort of 197 preoperative scans from patients who underwent hepatic resection for treatment of CRLM was used to evaluate the prognostic value of features and models to predict overall survival. A standard set of 93 features was extracted from all images, with a set of eight different extractor settings. The feature extraction settings producing the most reproducible, as well as the most prognostically discriminative feature values were highly dependent on both the region of interest and the specific feature in question. While the best overall predictive model was produced using features extracted with a particular setting, without accounting for reproducibility, (C-index = 0.630 (0.603--0.649)) an equivalent-performing model (C-index = 0.629 (0.605--0.645)) was produced by pooling features from all extraction settings, and thresholding features with low reproducibility ($\mathrm{CCC} \geq 0.85$), prior to feature selection. Our findings support a data-driven approach to feature extraction and selection, preferring the inclusion of many features, and narrowing feature selection based on reproducibility when relevant data is available.