3756 - Expanding the Role of PSMA-PET: A Novel Approach to Prostate Cancer Prognosis

Purpose/Objective(s): Predicting recurrence and metastasis in locally advanced prostate cancer following definitive radiation therapy remains a significant clinical challenge. Early and accurate risk stratification through advanced imaging is essential for guiding personalized treatment strategies and optimizing patient outcomes. While PSMA-PET imaging has revolutionized metastatic detection in prostate cancer, its primary clinical application has been in post-surgical settings, where the primary tumor has already been removed. Its potential as a prognostic tool for assessing the primary tumor before treatment remains largely unexplored. This study aims to expand the role of PSMA PET-based molecular imaging in pre-treatment settings beyond its traditional use. By leveraging radiomics, we investigate its ability to serve as an early biomarker of recurrence and metastatic progression before treatment begins.

Materials/Methods: We retrospectively analyzed 394 prostate cancer patients who underwent comprehensive clinical evaluation—including clinical staging, PSA testing, and biomedical imaging—prior to definitive radiation therapy. Among these, 169 patients underwent pre-treatment PSMA-PET. With a post-radiation follow-up extended until November 1, 2023. Right-censored time-to-PSA-relapse data were extracted from patient registry, yielding a total of 26 PSA relapse events recorded. Among occurred events, mean time-to-PSA-relapse is reported 1015 days (IQR: 490 - 1565). For 165 patients, GTVs were found and manually contoured on the PSMA-PET images. Radiomics features such as first-order and second-order statistics were extracted using open source software without renormalization. CT features were extracted in 2D using a bin width of 25 HU, and PSMA features in the bin width of 0.3 SUV. Data were divided into training and testing cohorts (75/25 split). A Random Forest Survival (RSF) model (scikit-survival v0.23) was employed for model training and testing.

Results: In the testing cohort (n = 46; 8 PSA relapse events), the trained RSF model demonstrated strong predictive performance, achieving a concordance index (C-index) of 0.85. This indicates that 85% of right-censored data pairs were correctly ranked based on their predicted risk of PSA relapse. These results suggest that the model provides reliable discrimination between high- and low-risk patients. Among events that occurred, the risk score increases with shorter time-to-PSA-relapse. In the supplementary analysis using only first-order radiomic features, the model yielded a lower C-index of 0.65, reflecting reduced predictive accuracy.

Conclusion: Our preliminary results suggest that PET/CT radiomic features derived from pre-treatment GTVs are promising prognostic biomarkers for predicting biochemical recurrence in prostate cancer patients undergoing radiation therapy. Future work will focus on incorporating external validation and extended follow-up time to further validate this predictive model.