Machine learning models combining ⁶⁸ Ga-PSMA PET radiomics and clinical features to predict overall survival for mCRPC patients treated with ¹⁷⁷ Lu-PSMA radioligand therapy.

Abstract

215 Background: Metastatic castration-resistant prostate cancer (mCRPC) remains incurable despite recent therapeutic advances. 177 Lu-PSMA radioligand therapy (RLT) provides meaningful clinical benefit in selected patients, yet approximately 30% show limited or no response. Although various baseline prognostic factors are well known, reliable pre-therapeutic predictive biomarkers are still lacking. Among emerging solutions, radiomics—by extracting quantitative imaging features—offers a novel approach to capture tumor heterogeneity and predict outcomes beyond standard visual assessment. To address this need, we developed and compared several predictive machine learning models integrating clinical, biochemical, and imaging data—including 68 Ga-PSMA PET radiomic features—to estimate overall survival in mCRPC patients treated with 177 Lu-PSMA RLT. Methods: This retrospective monocentric study included 102 mCRPC patients treated with 177 Lu-PSMA between January 2022 and March 2024. Five lesions per patient were segmented on their pre-therapeutic 68 Ga-PSMA PET/CT. Radiomic features were extracted following IBSI guidelines. Associations with OS were evaluated using Cox proportional hazards models. Multimodal machine learning (ML) models integrating clinical, genomic, biochemical, and imaging features (including radiomics) were developed to predict OS. Results: Median OS of the total cohort was 11.2 months. Intralesional homogeneity characterized by 2nd order radiomic feature “Normalized Inverse Difference Moment” was associated with shorter OS (p=0.025). Among clinical parameters, ECOG ≥2, liver metastases, and high tumor burden (≥20 metastases or superbone scan) were significantly correlated with poor OS (p<0.05). The model leveraging PET data alone achieved strong performance, with a 12-month AUC comparable to the clinical model (0.715 vs 0.766). The best-performing ML model achieved a 12-month AUC of 0.778 by combining 22 clinical, biochemical, genomic, and radiomic features. Conclusions: Machine learning models based on 68 Ga-PSMA PET data, including radiomic features, offer a non-invasive, standardized, and widely accessible approach to predict survival in mCRPC patients treated with 177 Lu-PSMA RLT. Imaging-only models achieved performance comparable to clinical–biochemical models, providing robust prognostic information even in the absence of extensive clinical or molecular data. This highlights the strong predictive potential of PSMA PET imaging for patient stratification in real-world settings. Furthermore, radiomic characterization of intratumoral heterogeneity adds independent prognostic value, supporting the integration of imaging-derived biomarkers into personalized treatment decision-making.