Impact of gamma model-based harmonization on PET tissue classification

Purpose: Positron Emission Tomography (PET) is an advanced clinical imaging modality within the field of nuclear medicine. Despite strict adherence to standardized imaging protocols, variations in image indices caused by various factors complicate the comparison of multicenter PET data. This study aims to investigate whether a Gamma model-based standardized uptake value (SUV) harmonization algorithm can improve the classification of imaging radiomics features across multicenter datasets. Methods: This study includes PET/CT and PET/MR image data of 251 patients from Guangdong Provincial People's Hospital without significant lesions in the liver, spleen, and bone regions. A Gamma modelbased harmonization algorithm is developed and evaluated for its impact on radiomic-based tissue classification. The effectiveness of the harmonization algorithm is assessed based on classification accuracy, in comparison with unharmonized and ComBat-harmonized data. Results: The study identifies 86 radiomic features spanning five categories from volumes of interest (VOIs) in three tissues. Our harmonization algorithm significantly improved tissue classification accuracy compared to unharmonized data. By utilizing the PCA-reduced 10-dimensional radiomic features, the algorithm achieved a mean accuracy of 94.5% and 93.0% in the training and validation sets, respectively, representing a substantial improvement over the unharmonized accuracies of 55.8% and 52.0%. Conclusion: These findings suggest that the Gamma model-based harmonization algorithm facilitates the investigation of imaging radiomics in multicenter data.