Radiomics Features from Different Prostatic Zones on ¹⁸F-PSMA-1007 PET/CT for Predicting Persistent PSA in Prostate...

Objectives: This study aims to explore the role of radiomics features (RFs) from prostate subregions, including the tumor microenvironment (TME), in predicting persistent PSA. Methods: In retrospective analysis, we segregated 354 patients with pathologically confirmed localized prostate cancer (PCa) into training, internal validation, and external validation cohorts. The prostate on ¹⁸F-prostate-specific membrane antigen (PSMA)-1007 positron emission tomography/computed tomography (PET/CT) was partitioned into three zones based on the maximum standardized uptake value (SUVmax) (zone-intra: 45-100% SUVmax; zone-peri: 20-45% SUVmax; zone-norm: 0-20% SUVmax). RFs from these zones were harnessed to develop five radiomics models [model-intra; model-peri; model-norm; model-ip; model-ipn]. Three optimal radiomics models were further integrated with the PSA model to construct combined models. Model performance was evaluated using the receiver operating characteristic (ROC) curves and the area under the curve (AUC). Results: Utilizing least absolute shrinkage and selection operator (LASSO) and logistic regression, five radiomics models were constructed, with model-ip, model-ipn, and model-intra showing superior performance [training cohort AUCs: 0.76 (0.68-0.83), 0.75 (0.68-0.83), 0.76 (0.68-0.83); internal validation cohort AUCs: 0.76 (0.65-0.88), 0.72 (0.57-0.86), 0.70 (0.55-0.86); external validation cohort AUCs: 0.70 (0.50-0.86), 0.55 (0.36-0.73), 0.53 (0.34-0.72)]. Notably, the combined model incorporating model-ip and the PSA model exhibited optimal performance [training cohort AUC: 0.78 (0.71-0.85); internal validation cohort AUC: 0.78 (0.67-0.90); external validation cohort AUC: 0.89 (0.72-0.98)]. Conclusions: The RFs in different subregions on ¹⁸F-PSMA-1007 PET/CT have varying effectiveness in predicting persistent PSA. A radiomics model that encompasses the 20-45% SUVmax and 45-100% SUVmax zones, when combined with the PSA model, enhances predictive accuracy.