Artificial intelligence–derived tumor volume from baseline MRI and postprostatectomy outcomes following neoadjuvant hormonal therapy for patients with high-risk prostate cancer.

330 Background: Neoadjuvant hormonal therapy (NHT) with androgen deprivation therapy and androgen receptor pathway inhibitor (ARPI) before radical prostatectomy (RP) has emerged as a potentially promising strategy for patients with high-risk prostate cancer (PCa), with several trials, including PROTEUS, which are investigating this approach. Early studies demonstrated that 20-25% of patients have a pathologic complete response (pCR) or minimal residual disease (MRD) after NHT with ARPI, which is highly prognostic for long-term outcomes. However, predictors of pCR/MRD remain poorly understood. We previously demonstrated that the total volume of intraprostatic tumor (V AI ) on prostate multiparametric MRI, determined using an artificial intelligence (AI) model, is independently prognostic in localized PCa treated with upfront RP or radiotherapy. Our objective was to examine the association between pre-NHT V AI and outcomes after RP. Methods: We identified 105 patients with high-risk PCa from 4 prospective trials of NHT with ARPI who had pre-NHT MRIs between 2009-2018. A deep learning model for segmenting the intraprostatic tumor was trained using nnU-Net v2 with a ResNet encoder with an independent cohort of 1776 patients with localized PCa and MRIs utilizing 5-fold cross-validation. An ensemble model was used to provide segmentations for held-out pre-NHT MRIs. We examined the association between V AI and likelihood of pCR/MRD (≤5mm residual tumor) using logistic regression, as well as associations between V AI and time to next therapy (TT) and metastasis-free survival (MFS) using univariable Cox regression. Results: 26 patients (24.7%) had pCR/MRD. Median pre-NHT V AI for patients with and without pCR/MRD were 0.87mL (interquartile range [IQR] 0.25-1.53) and 2.50mL (IQR 1.16-4.30), respectively. Increasing V AI was associated with decreased likelihood of pCR/MRD on multivariable analysis (adjusted odds ratio 0.97 per mL increase, 95% confidence interval [CI] 0.95-1.00, p=0.027). V AI had a greater area under the receiver operating characteristics curve than NCCN clinical risk groups (0.778 vs 0.511, p<0.001) for predicting pCR/MRD. After median follow-up of 6.3 years, the hazard ratios for associations between V AI and TT and MFS were 1.14 (95% CI 1.08-1.21, p<0.001) and 1.07 (95% CI 0.99-1.16, p=0.091), respectively. 6-year MFS rates were 94.4%, 87.0%, and 75.4% for V AI of 0-0.4, 0.5-1.9, and ≥2.0mL, respectively. Conclusions: A higher V AI on baseline MRI was associated with decreased likelihood of pCR/MRD and shorter time to next therapy after NHT and RP. If validated, AI-determined tumor volume has potential for becoming a clinically useful tool for the upfront identification of both patients most likely to benefit from NHT and RP, as well as those who should be considered for alternative treatment strategies.