Systemic Imaging Signatures from Normal Organs In PSMA-Negative Prostate Cancer
Abstract
Purpose
To determine whether radiomic features extracted from normal endocrine organs, when combined with clinical variables, can capture systemic imaging signatures predictive of clinical progression in patients with PSMA-negative prostate cancer.
Methods
In this retrospective study, 101 men with biochemically recurrent prostate cancer (median age 75 years; range 55–92) and PSMA-negative [¹⁸F]DCFPyL PET/CT scans were included. Radiomic features were extracted from the adrenal glands, thyroid, hypothalamus–pituitary complex, and testes following standardized preprocessing. Radiomic features were integrated with clinical variables and analyzed using three feature-selection strategies (SelectKBest, LASSO, Recursive Feature Elimination) and five machine-learning classifiers (Logistic Regression, Support Vector Machine, Random Forest, Gradient Boosting, and Histogram-Based Gradient Boosting). Model performance was evaluated using stratified cross-validation and assessed by AUC, accuracy, sensitivity, specificity, Brier score, and log loss.
Results
Across all modeling strategies, CT-derived radiomic features combined with clinical variables achieved the highest predictive performance (AUC 0.82–0.86), particularly for models incorporating features from the hypothalamus–pituitary complex and testes, with or without thyroid inclusion. PET-only radiomic models demonstrated moderate but consistent performance (AUC 0.75–0.83). Frequently selected texture features, including GLCM Autocorrelation and GLSZM Zone Entropy, reflected tissue heterogeneity and captured reproducible systemic textural patterns associated with disease progression. The inclusion of clinical variables consistently improved both model discrimination and calibration.
Conclusion
Radiomic analysis of normal endocrine organs reveals systemic imaging phenotypes that are predictive of clinical progression in PSMA-negative prostate cancer. Integrating multi-organ CT and PET radiomics with clinical data enhances predictive performance and interpretability, supporting a “microenvironment-at-a-distance” framework for systemic disease characterization beyond tumor-centric imaging.