Nonlinear stability phenotypes and collapse dynamics of multi-organ radiomics in total-body PET imaging: toward digital twin-oriented robustness assessment

OBJECTIVE: Radiomics analysis in PET offers a high-dimensional representation of tissue characteristics, yet its reliability is limited by sensitivity to acquisition perturbations, particularly dose reduction. This study aims to introduce a nonlinear stability phenotyping framework to characterize dose-response behavior of multi-organ radiomics and evaluate their potential relevance for future digital twin-oriented applications. Approach: A multi-center dataset of 550 total-body PET scans was analyzed. Dose reduction was simulated using list-mode subsampling (dose reduction factor - DRF = 2-100). A total of 107 radiomics features were extracted per organ. Stability was quantified using Intraclass Correlation Coefficient (ICC) and summarized as stability fractions. Nonlinear models were fitted to organ- and feature-level trajectories. Stability phenotypes and critical thresholds were derived, along with organ-level instability metrics. Main results: Radiomics stability exhibited strongly nonlinear, organ-dependent behavior. Feature-level analysis identified frequent sigmoid-associated collapse patterns with high descriptive model fits (R²>0.90). The liver showed the highest instability burden (30.8% unstable features), followed by esophagus and kidneys (18-20%). Texture features were the primary source of instability, whereas shape features remained robust. Organ-level trajectories demonstrated heterogeneous responses, with some organs maintaining high stability (0.80-0.95) across DRFs, while others showed sharp declines (down to 0.64). Linear (stable) behavior dominated at the organ level, though masking underlying nonlinear feature dynamics. Significance: Radiomics stability in total-body PET follows nonlinear, phenotype-driven dynamics with identifiable tipping points and organ-specific sensitivity. The proposed framework provides a behavior-based approach for evaluating radiomics robustness and supports the selection of reliable features and organs for robust quantitative imaging and may inform future digital twin-oriented applications.