Evaluating consistency of radiomic features derived from CT images: A cross‐center phantom study

TL;DR

This study tests whether radiomic features from CT scans are consistent across different hospitals in a Canadian province, using a lung phantom to assess data harmonization for multi-center research.

Contribution

The study introduces a low-cost lung phantom to evaluate radiomic feature consistency across multiple centers and CT scanners.

Findings

Radiomic features showed high consistency across centers (ICC > 0.941) with no significant differences detected.

52.5% of features met the threshold for consistency (CV ≤ 0.10), suggesting some standardization is needed before combining data.

The phantom proved durable for transport and useful for quality assurance in multi-institutional settings.

Abstract

To investigate the consistency of radiomic features extracted from computed tomography (CT) scans across CT radiotherapy simulators geographically spread across a Canadian province using a simplified lung radiomic phantom, and to determine whether it is appropriate to combine multicenter imaging data into a single dataset.

An inexpensive phantom was created using foam with a density similar to lung and a plastic vial insert filled with water. It was imaged at six provincial radiotherapy treatment centers using eight GE CT radiotherapy simulators and routine lung stereotactic ablative radiotherapy planning CT acquisition protocols. Radiomic features were extracted from regions of interest using Imaging Biomarker Explorer radiomics software and compared using Kruskal Wallis H tests, intraclass correlation coefficient (ICC), and coefficient of variation (CV).

Image acquisition parameters were similar across centers. At the population level, no significant inconsistencies between radiomic features originating from different centers or from within the same center were observed (Bonferroni‐corrected p > 0.05; ICC > 0.941). On average, 52.5% of features were considered consistent (CV ≤ 0.10).

The proposed phantom was transported across widespread centers without detectable damage and demonstrates potential for easy quality assurance checks on radiomic feature consistency within a multi‐institutional setting. Our analysis suggests that some features should be omitted or standardized before combining provincial imaging data into a harmonized lung radiotherapy dataset. These preliminary findings lay the groundwork for further investigation into provincial radiomic feature consistency and potential application to multicenter clinical studies. Owing to potential differences in imaging protocols, a consistency evaluation should be performed before undertaking radiomic analysis of data combined from different institutions.

Figures