1348933 - ComBat Harmonization for Myocardial Radiomic Features Sensitive to Imaging Parameters in Cardiac MR Imaging

Recent studies have shown the feasibility and improved performance of cardiac MRI radiomics analysis over conventional imaging. However, variability in imaging parameters may affect the reliability and reproducibility of myocardial radiomic features. In this study, we evaluate the impact of radiomics harmonization method (ComBat method) to remove the scanner effects and improve the cardiac MRI radiomics reproducibility.

Methods:

The study included 16 subjects (11 healthy and 5 patients) from publicly available database (Harvard dataverse). Left ventricle myocardium was manually segmented and radiomic features were extracted using PyRadiomics from original image and after applying 9 image filters. Radiomic features (1652 per sequence) were extracted from 5 sequences (Cine, T1 DB, T1 Map, T2 DB, T2 Map) using original image and after varying multiple imaging parameters (flip angle 1(FA1), flip angle 2 (FA2), low resolution (LowRes), high resolution (HighRes), parallel imaging (mSENSE/GRAPPA), slice thickness 1 (ST1), and slice thickness 2 (ST2). For each of the 5 sequences, the data across all the imaging parameters was aggregated into one dataset (yielding 5 total datasets). For each of the datasets, radiomic features with zero variance and/or extreme outlier values were removed.

 
Sensitivity values for each radiomic feature were calculated using Cohen's effect size looking at the difference between radiomic feature values under the "original" imaging parameter and values under all the other imaging parameters. Sensitivity values less than 0.2 were considered robust. Furthermore, for each radiomic feature, Friedman's test was used to analyze whether there was a difference in the radiomic feature values between all the imaging parameters. The P-Values from each of these tests were aggregated into a "differential feature ratio," defined as the proportion of p-values < 0.05 across all the tests performed. The full procedure described was run twice, once using the straight radiomic features data and another time after applying the ComBat method to the data. ComBat is a harmonization method that attempts to transform data so as to remove "batch" effects from explaining its variation. ComBat affected the sensitivity of the radiomic features to various imaging parameters.

Results:

Only 51.4% of features were robust across imaging parameters without comBat and 83.0% of features were robust after comBat harmonization. In all qualitative sequences and T2 mapping, radiomic features were most sensitive to changes in in-plane spatial resolution with T1 mapping features being most sensitive to flip angle variation (Table 1). Combat harmonization reduced the sensitivity of radiomic features across all sequences and imaging parameters (for example, differential feature ratio in cine sequence (pre combat 0.45) and after combat ( 0.247) (Table 1).

Conclusion: Overall, across all 5 sequences, applying ComBat appeared to decrease the sensitivity of the radiomic features to the imaging parameters.