18F-FDG PET radiomics is emerging as a promising tool to identify imaging biomarkers for quantifying intratumor heterogeneity in lung cancer. However, the robustness of PET radiomic features (RFs) is influenced by factors such as image reconstruction algorithms, tumor segmentation, and discretization. Although the impact of these factors on RFs has been explored, the specific influence of the advanced block sequential regularized expectation maximization (BSREM) reconstruction algorithm remains unclear. This study investigated the potential variations in PET RFs associated with different factors when using BSREM. Methods: Retrospective data of 18F-FDG PET from 120 lung cancer patients were reconstructed twice using advanced BSREM and conventional ordered-subset expectation maximization methods. For each reconstruction set, 3 tumor segmentations were performed, including manual, 40% threshold, and Nestle methods. Two discretization methods using absolute and relative settings were applied for each dataset before RF extraction. Stable and robust RFs were assessed by the coefficient of variance and intraclass correlation coefficient, respectively. Results: High instability was exhibited by 19%, 33%, and 36% of RFs, with a coefficient of variation of more than 20% for reconstruction, segmentation, and discretization, respectively. Conversely, 60%, 19%, and 35% of RFs demonstrated robustness against these factors, with an intraclass correlation coefficient of more than 0.90. The comparative evaluation revealed significantly greater robustness for most RF subtypes in BSREM than in ordered-subset expectation maximization under varying segmentation and discretization conditions (P < 0.05). Conclusion: The stability and robustness of PET RFs are enhanced if BSREM is applied rather than the conventional method. Study results suggest that the advanced reconstruction method could offer potential benefits in providing consistent PET-based radiomic analysis for improving diagnostic and prognostic value.
Read full abstract