PSMA PET/MR radiomics to improve postsurgical Gleason score prediction in prostate cancer

Ziel/Aim Gleason score (GS) is an important parameter for patient management in primary prostate cancer (PC), but biopsy-based GS (bGS) underestimate the more reliable postsurgical GS (psGS), obtained after radical prostatectomy (RP). In a previous work, we showed that radiomics from prostate-specific membrane antigen (PSMA)-targeted PET and Apparent Diffusion Coefficient (ADC) maps can predict psGS. In this work, we investigate the performance of our PET/MR radiomics model against bGS in the prediction of psGS.

Methodik/Methods Ga-68-PSMA PET/MR studies for primary staging of PC from a single scanner were retrospectively included. GS were obtained from biopsies and after RP. A support vector machine (SVM) model was trained with IBSI-compliant whole-prostate radiomics from PET and ADC studies, to predict psGS in three groups (G1: GS < 8, G2: GS = 8, G3: GS > 8). From a 6-fold cross-validation with 2:1 train vs validation samples, the best performing SVM model (highest balanced accuracy in the validation) was selected. To compare the predictions from our model to the use of bGS, overall accuracy (Acc), balanced accuracy (bAcc) and accuracy by classes (Acc_c) were reported.

Ergebnisse/Results The model was trained and validated with 101 patients (G1: 59 % (n = 60); G2: 23 % (n = 23); G3: 18 % (n = 18)), and 71 patients with both available psGS and bGS were selected for the comparison (G1: 62 % (n = 44); G2: 20 % (n = 14); G3: 18 % (n = 13)). Our model outperformed the bGS in predicting psGS overall (Acc: 77.7 % vs 73.1 %; bAcc: 82.4 % vs 75.2, respectively) and within each group (Acc_c: G1: 72.0 % vs 70.7 %; G2: 84.0 % vs 68.0 %; G3: 91.3 % vs 87.0 %, respectively).

Schlussfolgerungen/Conclusions The combined information from PSMA-PET and ADC map radiomics featured a better performance, compared to assuming the bGS, in the prediction of psGS by groups.

• Literatur/References

• This project is funded by the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 764458

Publication History

Article published online:
08 April 2021