Impact of dose reduction and multi-parametric harmonization on multi-organ radiomics robustness in total-body PET...

Radiomic features derived from positron emission tomography (PET) are sensitive to count statistics and acquisition protocol variability, thereby limiting quantitative reliability in low-dose and multi-center total-body PET imaging.

This study evaluated dose robustness of PET radiomics and assessed the impact of multi-parameter ComBat harmonization on feature reproducibility.

A multi-center total-body PET dataset comprising 550 patients from two institutions was retrospectively analyzed. Seven dose levels were simulated from list-mode data (Standard; dose reduction factors (DRF) of 2, 4, 10, 20, 50, and 100). Automatic deep learning-based segmentation was applied. A total of 107 Image Biomarker Standardization Initiative (IBSI)-compliant radiomic features were extracted from standardized uptake value (SUV)-normalized images. Feature-level ComBat harmonization was performed under five configurations (scanner manufacturer, pixel spacing, radiopharmaceutical, scatter correction, and non-harmonized baseline). Reproducibility was quantified using intraclass correlation coefficient (ICC (2,1)) across dose levels, organs, and harmonization settings.

Dose reduction produced organ- and feature-dependent effects. Overall, 32.7% of the features achieved good or excellent reproducibility (ICC ≥ 0.75) in at least one anatomical region. Selected first-order and intensity-associated features retained good reproducibility in several anatomical regions under moderate dose reduction (DRF ≤ 20), whereas severe count reduction (DRF 50-100) was associated with greater organ- and feature-dependent instability, particularly among higher-order texture features in vascular and lower-contrast soft-tissue regions. Harmonization produced organ and feature-dependent changes but did not uniformly improve cross-dose reproducibility or compensate for noise-induced instability.

Radiomics robustness in total-body PET is strongly influenced by dose level, anatomy, and acquisition variability. Moderate dose reduction preserves selected intensity-based features, while extreme count reduction compromises higher-order textures. Harmonization altered reproducibility patterns but did not uniformly improve cross-dose agreement or fully offset noise-induced instability, underscoring the need for dose-aware feature selection in multi-center PET radiomics.