ComBat Harmonization of Image Reconstruction Parameters to Improve the Repeatability of Radiomics Features

• Published in: IEEE conference record - Nuclear Science Symposium & Medical Imaging Conference. pp. 1 - 3
• Main Authors: Hajianfar, Ghasem, Avval, Atlas Haddadi, Sabouri, Maziar, Khateri, Maziar, Jenabi, Elnaz, Geramifar, Parham, Oveisi, Mehrdad, Shiri, Isaac, Zaidi, Habib
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.10.2021
• Subjects: Clinical trials; Feature extraction; Filtering algorithms; Imaging phantoms; Lesions; Reconstruction algorithms; Thorax
• ISSN: 2577-0829
• DOI: 10.1109/NSS/MIC44867.2021.9875440

Image reconstruction parameters lead to variability in radiomic features, which challenges repeatability and reproducibility of radiomics features, especially in multicenter clinical trials. To deal with this issue, a number of harmonization techniques were proposed. The aim of the current study was to investigate feature domain harmonization to cope with repeatability issues in PET radiomic features. Our study was conducted on a thorax phantom filled with 18F-FDG to acquire PET/CT images followed by extraction of 56 radiomics features on each of the six phantom lesions. Different reconstruction algorithms, post-processing filters, number of iterations, number of subsets, and matrix sizes were investigated. ComBat harmonization method was applied to these parameters on radiomics features and the results and associated p-values reported using Kruskal-Wallis test with a significance level of 0.05. This test indicated that 2, 25, 8, 26, and 29 features for reconstruction algorithms, post-processing filter, number of iterations, number of subset sand matrix size parameters, respectively, had significant variability (all p-values <0.05) before harmonization. These were reduced to 0, 2, 0, 0, and 0 features, respectively. The results of our study indicate that the repeatability of PET radiomics features among several image reconstruction parameters might be improved with the help of harmonization methods and could further support multi-institutional studies.