PSMA PET/CT radiomics: Assessment of Adverse Pathological Risk and Proteomic Biomarker Correlations in Prostate Cancer

Introduction: Prostate cancer (PCa) is a highly heterogeneous malignant disease. Therefore, it is crucial to investigate markers to facilitate early identification of adverse pathological features of PCa thereby improve patient prognosis. In this study, we applied radiomics machine learning models to predict the aggressiveness of PCa and identify quantitative radiomic features and protein biomarkers associated with poor pathological traits. The aim of the study was to construct a multi-omics marker model to optimize clinical risk stratification.

Methods: This was a retrospective study on 191 patients who were diagnosed with PCa or benign prostatic hyperplasia (BPH) and were pathologically confirmed after undergoing ⁶⁸Ga-PSMA-617 PET/CT scan. CT imaging was utilized for anatomical localization, while PET/CT scans were employed for image fusion and manual contouring of the prostate gland was performed. Radiomic features were then extracted from the contours to analyze the imaging characteristics. Six machine learning algorithms were applied to construct radiomics models for predicting malignancies and combinations of adverse pathological features (Gleason score (GS), ISUP group, pathological stage (pT), lymph node infiltration (LNI), and perineural invasion (PNI). Two methods, minimum redundancy maximum relevance (mRMR) and LASSO, were utilized conduct feature selection and identify quantitative radiomic features with high predictive ability. Moreover, proteomics analyses were performed on 39 patients to identify protein biomarkers associated with adverse pathological features at the molecular level in PCa. Correlation analysis was performed to determine the association of quantitative radiomic features with protein biomarkers.

Results: The optimal radiomics model constructed using machine learning methods showed an area under the curve (AUC) of 0.938 (95% CI: 0.893 to 0.983) for predicting malignant prostate lesions and an AUC of 0.916 (95% CI: 0.854 to 0.977) for adverse pathological feature combinations in the test set. Results of the validation set obtained AUC values of 0.918 (95% CI: 0.848 to 0.989) for predicting malignancy and 0.855 (95% CI: 0.728 to 0.983) for adverse feature combinations. Three quantitative radiomic features and ten protein molecules associated with adverse pathological characteristics were identified. Moreover, a significant correlation was observed between quantitative radiomic features and protein biomarkers. The radioproteomic analysis demonstrated that molecular changes in protein molecules could affect the imaging biomarkers.

Conclusions: This study underscored the efficacy of radiomics machine learning models using ⁶⁸Ga-PSMA-617 PET/CT in stratifying PCa risks. Our models demonstrated high predictive accuracy for malignancy and adverse pathological features, evidenced by robust AUC values. Notably, we identified critical quantitative radiomic features and protein biomarkers, revealing significant correlations between imaging and molecular changes in PCa. The integration of these findings into a multi-omics marker model marked a significant stride in optimizing clinical risk stratification, potentially enhancing personalized treatment strategies and improving patient outcomes in PCa management.

• Download figure
• Open in new tab
• Download powerpoint

• Download figure
• Open in new tab
• Download powerpoint