AutoComBat: a generic method for harmonizing MRI-based radiomic features.

Area of Science:

• Radiomics
• Medical Imaging
• Machine Learning

Background:

• Multicentric Magnetic Resonance Imaging (MRI) data are crucial for generalizable radiomic signatures in brain oncology.
• Scanner and acquisition heterogeneity significantly impacts quantitative radiomic features.
• Existing harmonization methods like ComBat have limitations, including the need for batch labels and representative samples.

Purpose of the Study:

• To compare a priori and a posteriori radiomic harmonization methods.
• To propose a machine learning-compatible code adaptation for harmonization.
• To develop AutoComBat for automatic batch label determination using MRI metadata or quality metrics.

Main Methods:

• Compared preprocessing and ComBat-based harmonization techniques.
• Developed AutoComBat utilizing constrained clustering with MRI metadata or quality metrics.
• Evaluated methods on a heterogeneous glioma dataset from eight centers.
• Assessed harmonization by decreasing feature standard deviation and classification task performance.

Main Results:

• ComBat and AutoComBat with quality metrics showed promising white matter harmonization, but results varied across MRI sequences.
• Preprocessing improved T1-weighted post-contrast (T1w-gd) images for the grading task.
• AutoComBat, using metadata alone or with quality metrics, outperformed conventional ComBat for T2-weighted fluid-attenuated inversion recovery (T2w-flair) harmonization.

Conclusions:

• Radiomic harmonization effectiveness is dependent on MRI weighting, feature class, and task.
• AutoComBat demonstrates potential for robust data harmonization in brain oncology.
• Further investigation on diverse datasets is warranted to validate these findings.