[18F]F-DCFPyL PET/MRI radiomics for intraprostatic prostate cancer detection and metastases prediction using whole-gland segmentation.

[OBJECTIVES] To evaluate [18F]F-DCFPyL PET/MRI whole-gland-derived radiomics for detecting clinically significant (cs) prostate cancer (PCa) within the prostate gland and predicting extra-prostatic metastasis (N and M staging).

[METHODS] In this single-centre, retrospective study, therapy-naïve PCa patients who underwent [18F]F-DCFPyL PET/MRI were included. Whole-prostate segmentation was performed. Feature extraction from each modality was done. The selection of potential variables was made through regularized binomial logistic regression. The oversampled training data were used to train binomial logistic regression for each outcome. The estimates of the models were calculated, and the mean accuracy was reported. The trained models were assessed on the test data for comparative evaluation of performance.

[RESULTS] A total of 103 patients (mean age = 65; mean PSA = 23.4) were studied. Among them, 89 had csPCa and 20 had metastatic disease. There were five radiomics variables selected for the International Society of Urological Pathology Grade Group (ISUP GG) ≥ 2 from T2w, ADC, and PET. To detect N1, five radiomics variables were selected from the T2w and PET. For M1, four radiomics variables were selected from T2w and ADC. Regarding the performance of models for the prediction of csPCa, the imaging-based hybrid model (T2w + PET) provided the highest AUC (0.98). The performance of N1 models showed the highest AUC (0.80) for T2w + PET. To predict M1, the T2w + ADC model showed the highest AUC (0.93).

[CONCLUSIONS] Whole-gland PET/MRI radiomics may provide a reliable model to predict csPCa. Also, acceptable performance was reached for predicting metastatic disease in our limited population. Our findings may support the value of whole-gland radiomics for non-invasive csPCa detection and prediction of metastatic disease.

[ADVANCES IN KNOWLEDGE] Whole-gland PET/MRI radiomics, a less operator-dependent segmentation method, can be potentially used for treatment personalization in PCa patients.

[TRIAL REGISTRATION] NCT03535831. Registered 2018; NCT03149861. Registered 2017.