Harmonization of Myocardial Perfusion SPECT Radiomics Features: A Patient Study

Introduction: Radiomics enables diagnostic and predictive models for different disease. Use of different reconstruction methods and parameters can lead to high variability in radiomics features, rendering them unreliable and unreproducible. To address this challenge, several harmonization approaches have been suggested such as ComBat harmonization which has the ability to reduce batch-effects. This work aims to investigate the impact of ComBat harmonization on radiomics features as extracted from Myocardial Perfusion Imaging (MPI) using SPECT with different reconstruction parameters.Objective: Radiomics enables diagnostic and predictive models for different disease. Use of different reconstruction methods and parameters can lead to high variability in radiomics features, rendering them unreliable and unreproducible. To address this challenge, several harmonization approaches have been suggested such as ComBat harmonization which has the ability to reduce batch-effects. This work aims to investigate the impact of ComBat harmonization on radiomics features as extracted from Myocardial Perfusion Imaging (MPI) using SPECT with different reconstruction parameters.

Methods: In this work, we analyzed 20 patients with stress and rest SPECT MPI (40 scans). Different reconstructions were studied by varying reconstruction method (FBP vs. iterative), filter (Gaussian, No filter, Butterworth), order (5, 10), cutoff (0.35, 0.40, 0.45, 0.50, 0.55), and iteration-subset ((4-4), (4,6), (4,8), (6,4), (6,6), (8,4), (8,6)). Prior to feature extraction, isotropic resampling to 6.591×6.591×6.591 mm³ was performed. Intensity in the volume of interest (VOI) was normalized to mean±3´SD, followed by intensity discretized to 64 fixed number of bins. First-order and textural features were extracted from the 600 reconstructed images, and 43 IBSI based radiomic features were obtained from each image. To deal with variation in radiomics features, ComBat harmonization was applied on different kinds of reconstruction parameters. Finally, the Kruskal-Wallis test was performed on the data before and after ComBat harmonization, and p-values were reported (p<0.05 as significant).

Results: Given the Kruskal-Wallis test, the number of features before ComBat harmonization which had significant differences over reconstruction method, filter, order, cutoff, and iteration-subset were 11, 10, 0, 21, and 1, respectively. However, following ComBat harmonization, all these numbers dropped to 0 and no significant differences remained.

Conclusions: Several variations such as changes in reconstruction methods and parameters can significantly impact images, leading to variability in radiomics features, and under such situations, radiomic features are not reproducible anymore. In this study, we aimed to tackle this challenge by using harmonization. Applying ComBat harmonization to the data could sufficiently tackle this issue, providing a solution towards improved reproducibility of images and radiomics features generated from different reconstruction methods and parameters.

• Download figure
• Open in new tab
• Download powerpoint