End-to-end [F]PSMA-1007 PET/CT radiomics-based pipeline for predicting ISUP grade group in prostate cancer.

OBJECTIVES

To develop an end-to-end radiomics-based pipeline for the prediction of International Society of Urological Pathology grade group (ISUP GG) in prostate cancer (PCa).

METHODS

This retrospective study includes 356 patients (241 in training set and 115 in independent test set) with histopathologically confirmed PCa who underwent [F]PSMA-1007 PET/CT scan. Patients were classified into two groups according to their ISUP GG (1-3 vs. 4-5). Radiomics features were extracted from the whole, automatically segmented prostate on PET/CT images, 30 models were constructed by combining 6 feature selection algorithms and 5 machine learning classifiers. The clinical model incorporated age, total prostate-specific antigen (tPSA), maximum standardized uptake value (SUVmax), and prostate volume. The predictive performance of the models was evaluated using the area under the receiver operating characteristic curve (AUC), balanced accuracy (bAcc), and decision curve analysis (DCA).

RESULTS

The best-performing radiomics model significantly outperformed clinical model (AUC 0.879 ± 0.041 vs. 0.799 ± 0.051, bAcc 0.745 ± 0.074 vs. 0.629 ± 0.045). On an external independent test set, best-performing radiomics model perform better than clinical model, with an AUC of 0.861 vs. 0.750, p = 0.002 (Delong), and bAcc of 0.764 vs. 0.582, p = 0.043 (McNemar). The learning curve, calibration curve and DCA demonstrated goodness-of-fit and improved benefits in clinical practice.

CONCLUSION

The end-to-end radiomics-based pipeline is an effective non-invasive tool to predict ISUP GG in PCa.