MP15-02: Deep learning using MRI and intraoperative video images strongly predicts recovery of urinary continence after robot-assisted radical prostatectomy

Makoto Sumitomo*, Atsushi Teramoto, Wataru Nakamura, Kenji Zennami, Takuhisa Nukaya, Masashi Takenaka, Kosuke Fukaya, Manabu Ichino, Kiyoshi Takahara, Hitomi Sasaki, Ryoichi Shiroki, Toyoake, Japan

Introduction: We have recently reported that deep learning (DL) using pelvic magnetic resonance imaging (MRI) is useful for predicting the severity of urinary incontinence (UI) after robot-assisted radical prostatectomy (RARP). However, our results showed limitations since the prediction accuracy remained around 70% (Sumitomo M et al, Int J Urol, 2020). We aimed to make a more accurate prediction system that can be used to inform patients of the accuracy of recovery of UI after RARP using DL model from intraoperative video images in addition to MRI information.

Methods: This study included 102 patients with prostate cancer (PC) who underwent RARP. Patients using 0 or 1 pad/day within 6 months after RARP were categorized into the “good” group, whereas the other patients were categorized into the “bad” group. Three snapshots showing pelvic cavity (before bladder neck incision, just after prostate removal, after vesicourethral anastomosis) from intraoperative video records in addition to MRI DICOM data, and preoperative and intraoperative covariates were assessed. To evaluate the DL models from the testing dataset, their sensitivity, specificity and area under the receiver operating characteristic curve (AUC) were analyzed.

Results: Eighty-two patients (80%) were shown to  belong to the good continence group and 20 (20%) to the bad continence group at the six-month follow-up. The combination of DL and repeated machine learning (ML) using axial MRI and three intraoperative snapshots with or without clinicopathological parameters had the highest performance with sensitivity of 75.0%, specificity of 87.5%, and AUC of 81.3% for predicting recovery from UI after RARP, whereas DL and ML using axial MRI only achieved lower performance with sensitivity of 72.0%, specificity of 67.5%, and AUC of 72.1%.

Conclusions: Our results suggest that DL algorithms using preoperative MRI and intraoperative video images can be used to inform patients of the accuracy of recovery of UI after RARP. Further studies will determine what kind of parameters from intraoperative video images DL focuses for the classification.

Source of Funding: None