Improved generalized ComBat methods for harmonization of radiomic features.

Area of Science:

• Medical Imaging
• Radiomics
• Precision Medicine

Background:

• Radiomic analysis in precision medicine faces challenges due to data heterogeneity from varying image acquisition.
• Multicenter datasets often exhibit multimodal feature distributions, impacting generalizability.
• Existing harmonization tools like ComBat have limitations in handling multiple batch effects and non-normally distributed data.

Purpose of the Study:

• To develop and evaluate novel methods for harmonizing heterogeneous radiomic datasets.
• To address challenges of multiple batch effects and data bimodality in radiomic feature distributions.
• To assess the impact of harmonization on the predictive performance of radiomic models.

Main Methods:

• Proposed OPNested ComBat for sequential harmonization of multiple batch effects in an optimized order.
• Introduced Gaussian Mixture Model (GMM) grouping to address data bimodality, integrated as a batch variable (OPNested + GMM) or clinical covariate (OPNested - GMM).
• Evaluated methods using radiomic features from two public lung CT datasets, extracted with CapTK and PyRadiomics.

Main Results:

• OPNested ComBat demonstrated superior harmonization performance compared to standard ComBat.
• OPNested + GMM ComBat achieved the best harmonization but lowest predictive performance.
• OPNested - GMM ComBat showed reduced harmonization but improved predictive performance.

Conclusions:

• Improved harmonization does not always guarantee enhanced predictive performance in radiomic studies.
• The proposed OPNested ComBat and GMM-integrated methods show promise for standardizing heterogeneous datasets.
• These techniques can potentially improve the generalizability of radiomic findings across different imaging parameters and centers.