Supervised vs. unsupervised radiomics reproducibility across dose-level variations and partial volume correction methods in total-body PET imaging

Background and objective : To provide a systematic reproducibility analysis of radiomic features in total-body [ 18 F]-FDG PET imaging, assessing the effects of dose variation, partial volume correction (PVC), and harmonization. Methods : A dual-center total-body [ 18 F]-FDG PET dataset (n = 170 scans, 85 per center) acquired on the uEXPLORER and Biograph Vision Quadra systems was retrospectively subsampled from list-mode data to simulate four relative count levels (100%, 50%, 10%, 2%). PVC was performed using Richardson–Lucy (RL) and reblurred Van-Cittert (RVC) deconvolution, with non-PVC images serving as reference. Radiomic features (n = 93) were extracted from PET images. ComBat harmonization was applied across centers. Reproducibility was assessed using: (i) intraclass correlation coefficient (ICC), (ii) PCA/UMAP-based drift distances, and (iii) Linear Discriminant Analysis (LDA). Results : Across-dose ICC showed that > 60% of features fell into the Poor category (ICC < 0.5), with only 1–3 features (<3%) achieving Excellent reproducibility (ICC ≥ 0.9). Drift distances increased with dose reduction, from mean 5.98 at 50% to > 11 at 2%, reflecting strong dose dependence. PVC offered limited benefit: RL slightly outperformed non-PVC, whereas RVC consistently induced the largest drifts (up to 61% higher). ComBat harmonization failed to improve stability and occasionally increased variability (+13% for RL at 50%). LDA confirmed these findings: dose levels showed measurable separability in discriminant space, whereas PVC methods overlapped substantially, and harmonization effects were negligible, with before–after conditions appearing nearly indistinguishable. Conclusions: Radiomics reproducibility in total-body PET is primarily driven by dose level, with PVC effects being task-dependent and harmonization providing negligible benefit.