Deep learning using preoperative MRI information to predict early recovery of urinary continence after robot-assisted radical prostatectomy.

Abstract

310 Background: Urinary incontinence remains one of the most bothersome postoperative complications even after robot-assisted radical prostatectomy (RARP). We aimed to make a novel prediction system that can be used preoperatively to inform patients of the accuracy of early recovery of urinary continence after RARP using a deep learning (DL) model from magnetic resonance imaging (MRI) information and preoperative clinicopathological parameters. Methods: A retrospective cohort study was conducted on prostate cancer (PC) patients who had undergone RARP at our hospital between August 2015 and July 2019. Patients using no pads/no leakage of urine or the use of a safety pad within 3 months after RARP is categorized into “good” continence and others into “no good” continence. MRI DICOM data from axial, coronal and sagittal imaging as well as preoperative clinicopathological covariates (age, BMI, prostate volume, serum PSA level, Gleason score, clinical stage) were assessed. Supervised DL algorithms, which included AdaBoost, Naive Bayes, Neural Network, Random Forest, and SVM were trained and tested as binary classifiers (good or no good). To evaluate the DL models from the testing data set, their sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), as well as area under the receiver operating characteristic curve (AUC) were analyzed. Results: Data were available for 497 patients in the study period. The AdaBoost DL algorithm using MRI information in addition to clinicopathological parameters had the highest performance with sensitivity at 92%, specificity at 77%, PPV at 79%, NPV at 91%, and AUC at 84% for predicting good continence, while that using clinicopathological parameters only had the performance with sensitivity at 50%, specificity at 69%, PPV at 60%, NPV at 60%, and AUC at 60%. Conclusions: Our results suggest that the DL algorithms using MRI imaging information are highlighted as an accurate method for strongly predicting early recovery of urinary continence after RARP. Thus, DL predictions may help allocation of treatment strategies for PC patients who dislike prolonged urinary incontinence after RARP.