Indications for Adjuvant Radiation after Radical Prostatectomy as Predicted by Artificial Intelligence-Derived Dominant Intraprostatic Lesion Volume.

Abstract

In prostate cancer, PI-RADs scores of dominant intraprostatic lesions (DILs) in multi-parametric magnetic resonance imaging (mpMRI) are prognostic; however, their inter-observer agreement is only moderate. Artificial intelligence (AI) may be a powerful tool for prognostication by analyzing a large number of scans consistently in a short amount of time. This study investigated whether the DIL volume (DILvol) provided by an AI deep-learning segmentation algorithm could predict adverse findings at radical prostatectomy (RP), some of which could warrant adjuvant radiation therapy (RT). We conducted a retrospective study of 185 consecutive patients with localized prostate cancer who underwent an endorectal coil, high B-value (> = 1000 s/mm2), 3-Tesla mpMRI followed by RP between 2015 and 2017. Using a previously trained deep learning nnUNet algorithm for providing DIL segmentations from patients treated with definitive RT, we segmented the DIL for the RP cohort. We evaluated the association of AI DILvol with the risks of adverse pathologic factors, including positive margins, pathologic T3 (pT3) disease, and pathologic Gleason (pGS8-10) disease, using separate univariate logistic regression models. We then included AI DILvol, pT3 (vs pT2), pGS8-10 (vs pGS6-7), margin status, and pre-RP PSA for predicting post-RP PSA values utilizing multivariate linear regression analysis. Finally, we included these same factors into a multivariate logistic regression analysis for predicting the risk of meeting adjuvant RT indications (PSA persistence post-RP > = 0.1 ng/mL or positive lymph nodes). The median time between RP and post-PSA value was 1.6 months. The Pearson's correlation coefficient between AI and reference DILvol (sum of manually contoured PI-RADS 3-5 lesions) was 0.86 (p < 0.001). The Pearson's correlation coefficient between AI DILvol and pathologic tumor size was 0.63 (p < 0.001). Utilizing separate univariate logistic regression models, we found that AI DILvol was significantly associated with the risks of positive margins (OR 1.31 [1.10, 1.58]; p = 0.003), pT3 (OR 1.59 [95% CI: 1.30, 1.99]; p < 0.001), and pGS8-10 (OR 1.28 [1.07, 1.56]; p = 0.01). On multivariate linear regression, AI DILvol (0.27/mL [0.25, 0.29]; p < 0.001) was significantly correlated with post-RP PSA values, after controlling for adverse factors and pre-RP PSA. On multivariate logistic regression, AI DILvol (adjusted OR 1.32 [1.05, 1.69]; p = 0.03) was the only factor significantly associated with the risk of meeting adjuvant RT indications after controlling for these same factors. For localized prostate cancer treated with RP, AI DILvol was the only factor significantly associated with the risk of meeting adjuvant RT indications, even after controlling for pathologic factors at RP. Further studies are needed to determine if AI DILvol is prognostic for long-term oncologic outcomes after RP.