Related Experiment Video
Updated: Dec 17, 2025
10:17
Guidelines and Experience Using Imaging Biomarker Explorer IBEX for Radiomics
Published on: January 8, 2018
13.6K
Performance comparison of modified ComBat for harmonization of radiomic features for multicenter studies
R Da-Ano1, I Masson2,3, F Lucia2,4
1INSERM, UMR 1101, LaTIM, University of Brest, Brest, France. ronrickarnaiz@gmail.com.
Scientific Reports
|June 26, 2020
Summary
Statistical harmonization methods improve radiomic models for predicting patient outcomes. Modified ComBat versions enhance flexibility and robustness, leading to better predictive accuracy in multicenter studies for cervical and laryngeal cancers.
Area of Science:
- Medical imaging analysis
- Radiomics and computational pathology
- Statistical modeling in healthcare
Background:
- Radiomics offers prognostic value but faces challenges due to data variability across institutions.
- The ComBat method addresses "center-effects" in radiomics, but requires enhancements for flexibility and robustness.
- Standardization is crucial for pooling multicenter radiomics data for robust clinical applications.
Purpose of the Study:
- To evaluate modified ComBat versions (M-ComBat, B-ComBat, BM-ComBat) for harmonizing radiomic features in multicenter settings.
- To assess the impact of these harmonization techniques on the predictive performance of machine learning models.
- To determine if enhanced ComBat versions improve robustness and flexibility in radiomics data analysis.
Main Methods:
- Two modified ComBat versions were developed: M-ComBat (reference flexibility) and B-ComBat (bootstrap/Monte Carlo robustness).
- BM-ComBat combined both modifications. Four versions (original ComBat, M-ComBat, B-ComBat, BM-ComBat) were tested on two datasets: cervical cancer (MRI/PET) and laryngeal cancer (CT).
- Unsupervised clustering was used to define labels in the laryngeal cancer dataset due to heterogeneity.
Main Results:
- All ComBat versions successfully removed significant distribution differences in radiomic features between centers.
- Harmonization consistently improved the predictive ability of radiomic models across all tested machine learning pipelines and metrics.
- The improved ComBat versions (M-ComBat, B-ComBat, BM-ComBat) yielded superior results compared to the original ComBat.
Conclusions:
- Modified ComBat versions offer greater flexibility and robustness for harmonizing radiomics data in multicenter studies.
- These enhanced methods consistently improve the predictive power of radiomic models for clinical outcomes.
- The proposed modifications are essential for advancing the clinical utility of radiomics as a prognostic tool.
