Radiomics-based machine learning models for predicting genomic alterations in metastatic prostate cancer using PSMA PET imaging: a pilot study.
[BACKGROUND AND OBJECTIVE] Genomic characterization of metastatic prostate cancer (mPCa) plays a pivotal role in guiding precision oncology.
APA
Scavuzzo A, Pasini G, et al. (2025). Radiomics-based machine learning models for predicting genomic alterations in metastatic prostate cancer using PSMA PET imaging: a pilot study.. EJNMMI reports, 9(1), 42. https://doi.org/10.1186/s41824-025-00280-6
MLA
Scavuzzo A, et al.. "Radiomics-based machine learning models for predicting genomic alterations in metastatic prostate cancer using PSMA PET imaging: a pilot study.." EJNMMI reports, vol. 9, no. 1, 2025, pp. 42.
PMID
41467936
Abstract
[BACKGROUND AND OBJECTIVE] Genomic characterization of metastatic prostate cancer (mPCa) plays a pivotal role in guiding precision oncology. This study aimed to evaluate the feasibility of combining radiomics and clinical data within a machine learning (ML) framework to non-invasively predict key genomic mutations in patients with mPCa undergoing PSMA PET imaging.
[METHODS] A retrospective cohort of 14 mPCa patients who underwent [ F]PSMA-1007 PET/CT was analysed. Prostate and metastatic lesions were segmented, and radiomics features were extracted. Somatic genomic alterations were obtained from formalin-fixed paraffin-embedded tissue samples using FoundationOne CDx testing. Six ML algorithms - Discriminant Analysis, Support Vector Machines, K-Nearest Neighbours, Neural Networks, Random Forest, and Boosting - were trained using a 5-times repeated pipeline with 80/20 train/test split, LASSO feature selection, and 5-fold cross-validation. Model performance was assessed using accuracy, AUC, sensitivity, specificity, precision, and F-score.
[KEY FINDINGS] Fourteen patients with mPCa were included, and 46 lesions were analysed. Genomic alterations included mutations in TP53, TMPRSS2, PTEN, BRCA1/2, ATM, and others. Owing to data limitations, mutations other than TP53, TMPRSS2, and PTEN were grouped into a composite "OTHER" category. The best-performing clinical-radiomics ML models achieved AUCs of 91.11% (TP53), 84.44% (TMPRSS2), 80.00% (PTEN), and 77.78% (OTHER). Selected feature stability was consistent across repeated runs.
[CONCLUSIONS AND CLINICAL IMPLICATIONS] Clinical-radiomics ML models based on PSMA PET imaging show promising accuracy in predicting actionable genomic alterations in mPCa. These findings support further investigation into radiogenomics modelling as a complementary, non-invasive tool to inform molecular profiling and treatment stratification.
[METHODS] A retrospective cohort of 14 mPCa patients who underwent [ F]PSMA-1007 PET/CT was analysed. Prostate and metastatic lesions were segmented, and radiomics features were extracted. Somatic genomic alterations were obtained from formalin-fixed paraffin-embedded tissue samples using FoundationOne CDx testing. Six ML algorithms - Discriminant Analysis, Support Vector Machines, K-Nearest Neighbours, Neural Networks, Random Forest, and Boosting - were trained using a 5-times repeated pipeline with 80/20 train/test split, LASSO feature selection, and 5-fold cross-validation. Model performance was assessed using accuracy, AUC, sensitivity, specificity, precision, and F-score.
[KEY FINDINGS] Fourteen patients with mPCa were included, and 46 lesions were analysed. Genomic alterations included mutations in TP53, TMPRSS2, PTEN, BRCA1/2, ATM, and others. Owing to data limitations, mutations other than TP53, TMPRSS2, and PTEN were grouped into a composite "OTHER" category. The best-performing clinical-radiomics ML models achieved AUCs of 91.11% (TP53), 84.44% (TMPRSS2), 80.00% (PTEN), and 77.78% (OTHER). Selected feature stability was consistent across repeated runs.
[CONCLUSIONS AND CLINICAL IMPLICATIONS] Clinical-radiomics ML models based on PSMA PET imaging show promising accuracy in predicting actionable genomic alterations in mPCa. These findings support further investigation into radiogenomics modelling as a complementary, non-invasive tool to inform molecular profiling and treatment stratification.