The role of [68Ga]Ga-PSMA PET/CT in primary staging of newly diagnosed prostate cancer: predictive value of PET-derived parameters for risk stratification through machine learning

This study aimed to investigate the PSMA-avid distribution of disease in newly diagnosed prostate cancer (PC) and the correlation between [68Ga]Ga-PSMA-11 PET-derived parameters with serum PSA levels, biopsy Gleason Score (GS), and the presence of metastasis. Additionally, we explored whether machine learning-based analysis of PET-derived parameters predicts PSA value and biopsy GS.

We retrospectively evaluated 256 newly diagnosed PC patients who had undergone [68Ga]Ga-PSMA-11 PET/CT for staging after biopsy. Several primary tumors and whole-body SUV and volumetric parameters were extracted from PET images. The relationship between PSA value, GS, and metastatic tendency with PET-derived parameters was evaluated. Several classifiers were trained with PET-derived parameters to predict GS > 7 and PSA > 20.

Of the 256 evaluated patients, only seven cases (2.7%) showed a negative scan. Out of 249 positive cases, 137 (55%) exhibited only localized disease, while 112 (45%) showed signs of metastasis. There was a significant correlation between GS and PSA value with all PET-derived parameters related to the primary tumor (P < 0.05). In patients with metastatic scans, PET-derived parameters in the primary tumor were significantly higher compared to patients with only local disease (P < 0.05). Based on ROC curve analysis with AUC, the optimal PSA cut-off for a metastatic scan was 16.79 ng/ml. Furthermore, the optimal cut-off values for SUVmean, SUVmax, PSMA-TV, and TL-PSMA in the primary tumor for a metastatic c scan were 4.4, 12.99, 18.91, and 98.69, respectively. TL-PSMA demonstrated the highest AUC to predict GS ≤ 7 vs. >7 with an optimal cut-off of 75.37 cm³ and a sensitivity of 86% and specificity of 65%. Likewise, in the metastatic scans, wbTL-PSMA exhibited the highest AUC to predict GS ≤ 7 vs. >7 with an optimal cut-off of 106.60 cm³ and a sensitivity of 92% and specificity of 59%. TL-PSMA showed the highest AUC to predict PSA ≤ 20 vs. PSA > 20 with an optimal cut-off of 70.31 cm³ and a sensitivity of 81% and specificity of 66%. Additionally, in the metastatic scans, wbPSMA-TV demonstrated the highest AUC to predict PSA ≤ 20 vs. PSA > 20 with an optimal cut-off of 59.46 cm³ and a sensitivity of 76% and specificity of 63%. Among evaluated classifiers, linear support vector classifier (SVC), calibrated classifier CV and logistic regression demonstrated the highest accuracy for categorization of GS ≤ 7 and GS > 7. Furthermore, calibrated classifier CV, nearest centroid, and logistic regression showed the optimal accuracy in predicting PSA ≤ 20 and PSA > 20.

In conclusion, [68Ga]PSMA PET/CT is a valuable tool for evaluating primary PC, detecting lymph node spread and bone metastases. There is a correlation between GS and PSA value with PET-derived parameters, which can predict GS and metastatic potential. Lastly, utilizing machine learning to analyze PET-derived parameters can aid in predicting PSA value and GS in primary PC. These findings indicate a possible connection between the distribution and amount of PSMA expression detected on [68Ga]Ga-PSMA PET scans with both biopsy GS and PSA level.