Toward standardization: Assessing the reproducibility of radiomics features in partial volume-corrected brain PET images
Mohammad-Saber Azimi, Maryam Cheraghi, Fatemeh MahdiMaleki, Faezeh MahdiMaleki, Amirhossein Sanaat, Poul Flemming Høilund‐Carlsen, Abass Alavi, Habib Zaidi
Abstract
OBJECTIVE: This study aims to evaluate the reproducibility of radiomic features in brain positron emission tomography (PET) imaging across different brain regions and partial volume correction (PVC) methods, and to identify optimal feature classes and correction strategies for reliable clinical modeling. METHODS: This study analyzed 76 hybrid brain PET/MR images. Radiomic features were extracted from 21 anatomically segmented brain regions under seven conditions: uncorrected PET and six PVC techniques, including reblurred Van Cittert (RVC), Richardson-Lucy (RL), region-based voxel-wise (RBV), iterative Yang (IY), multi-target correction (MTC), and parallel level set (PLS) methods. A total of 93 features spanning six radiomics classes-First Order, Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), Gray Level Dependence Matrix (GLDM), and Neighborhood Gray Tone Difference Matrix (NGTDM)-were extracted using the PyRadiomics library. Reproducibility was assessed using the intraclass correlation coefficient (ICC) and coefficient of variation (COV). RESULTS: RVC and RL showed the best reproducibility, with over 60% of features having COV < 25% and ICC ≥ 0.75. In contrast, MTC and PLS resulted in the highest variability. GLCM and GLDM features were the most stable, while first order and NGTDM were the most variable. Regions, such as the cerebellum and lingual gyrus had the highest ICC values (≥ 0.9), whereas the fusiform gyrus and brainstem showed poor reproducibility (ICC < 0.5). CONCLUSIONS: Radiomics reproducibility in brain PET imaging is highly dependent on both the PVC method and anatomical region. RVC and RL are recommended for reliable quantitative analysis, particularly when used with robust feature classes, such as GLCM and GLDM. These findings emphasize the importance of methodological standardization and anatomically informed region-of-interest selection in radiomics research and clinical applications.