Association between Artificial Intelligence-Derived Tumor Volume and Oncologic Outcomes for Localized Prostate Cancer Treated with Radiation Therapy.

Abstract

Although clinical features of multi-parametric magnetic resonance imaging (mpMRI) have been associated with biochemical recurrence in localized prostate cancer, such features are subject to inter-observer variability. We evaluated whether the volume of the dominant intraprostatic lesion (DIL), as provided by a deep learning segmentation algorithm, could provide prognostic information for patients treated with definitive radiation therapy (RT). We conducted a retrospective study of 438 patients with localized prostate cancer who underwent an endorectal coil, high B-value, 3-Tesla mpMRI and were treated with definitive RT at our institution between 2010 and 2017. We utilized the publicly available nnUNet to train a segmentation model which was used to identify the DIL. We examined the association between the artificial intelligence (AI)-generated DIL volume and oncologic outcomes, including biochemical recurrence and metastasis risk, using cause-specific Cox regression and time-dependent receiver operating characteristic analysis. The AI model identified DILs with an area under the receiver operating characteristic (AUROC) of 0.827 at the patient level. For the 233 patients with available PI-RADS scores, with a median follow-up of 5.6 years, there were 28 biochemical failures. AI-defined DIL volume was significantly associated with biochemical failure (adjusted hazard ratio 1.60, 95% confidence interval 1.14-2.24, p = 0.007) after adjustment for PI-RADS score. Among all 438 patients with a median follow-up of 6.9 years, there were 49 biochemical failures and 22 metastases. The AUROC for predicting 7-year biochemical failure for AI volume (0.790) was similar to that for National Comprehensive Cancer Network (NCCN) category (p = 0.17). The AUROC for predicting 7-year metastasis for AI volume trended towards being higher compared to NCCN category (0.854 vs 0.769, p = 0.06). An AI algorithm using deep learning could identify the DIL with good performance. AI-defined DIL volume may be able to provide prognostic information independent of the NCCN risk group or other radiologic factors for patients with localized prostate cancer treated with RT.