Room: Stars at Night Ballroom 2-3
Purpose: To evaluate the ComBat harmonization algorithm for use in radiomics studies to reduce the variation in radiomics features from different imaging protocols.
Methods: The Gammex computed tomography (CT) electron density phantom and Quasar body phantom were imaged using different chest imaging protocols on 2 Siemens CT scanners at 6 reconstruction filters, 3 bowtie filters, 4 kVp levels, 3 slice thicknesses, and 2 pitches. Where possible, only one imaging parameter was changed for each variation. 105 radiomics features were extracted from 15 spatially varying spherical contours between 15mm and 20mm in each of the Lung300 density, Lung450 density, and wood inserts. The Kolmogorov-Smirnov test was used to determine significant differences in the distribution of the features from each protocol variation class (kVp, pitch, etc.) before and after ComBat harmonization. P-values were corrected for multiple comparisons using the Benjamini-Hochberg-Yekutieli procedure. Finally, the ComBat algorithm was applied to human subject data using 2 different thorax imaging protocols with 16 and 17 unique subjects that varied in bowtie filter and pitch. Spherical contours of lung and vertebral bone, 20mm and 10mm respectively, were used for radiomics feature extraction.
Results: ComBat harmonization reduced the percentage of significantly different features from 4%- 85% to 0-4%, from 2%- 86% to 0%-5%, and from 0%-78% to 0-10% across all protocol variations for the Lung450, Lung300 and wood inserts, respectively. For the human subject data, ComBat harmonization reduced the percentage of significantly different features from 17% and 26% for the bone and lung, respectively, to 0%.
Conclusion: ComBat harmonization is an effective means of harmonizing radiomics features extracted from different imaging protocols to allow comparisons in large multi-institution datasets. Biological variation can be preserved by providing the ComBat algorithm with a linear model of variables to protect. ComBat harmonization should be tested for its effect on predictive models.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by a grant from Varian Medical Systems. The authors disclose the following funding sources: Varian Medical Systems (RNM, GDH, EW), NIH (GDH, EW) and ViewRay (GDH). In addition, GDH has a partnership with Cardialogica and EW received royalties from UpToDate. MG has no disclosures.