Recognition of Postoperative Cystography Features by Artificial Intelligence to Predict Recovery from Postprostatectomy Urinary Incontinence: A Rapid and Easy Way to Predict Functional Outcome

Abstract

Purpose: Post-operative cystography has been used to predict the recovery of postprostatectomy urinary incontinence (PPI) in patients with localized prostate cancer. This study aimed to validate the predictive value of cystography for PPI and utilize a deep learning model to identify favorable and unfavorable features. Methods: Medical records and cystography images of patients who underwent robotic-assisted radical prostatectomy for localized prostate cancer were retrospectively reviewed. Specific cystography features, including anastomosis leakage, a downward bladder neck (BN), and the bladder neck angle, were analyzed for the prediction of PPI recovery. Favorable and unfavorable patterns were categorized based on the three cystography features. The deep learning model used for transfer learning was ResNet 50 and weights were trained on ImageNet. We used 5-fold cross-validation to reduce bias. After each fold, we used a test set to confirm the model’s performance. Result: A total of 170 consecutive patients were included; 31.2% experienced immediate urinary continence after surgery, while 93.5% achieved a pad-free status and 6.5% were still incontinent in the 24 weeks after surgery. We divided patients into a fast recovery group (≤4 weeks) and a slow recovery group (>4 weeks). Compared with the slow recovery group, the fast recovery group had a significantly lower anastomosis leakage rate, less of a downward bladder neck, and a larger bladder neck angle. Test data used to evaluate the model’s performance demonstrated an average 5-fold accuracy, sensitivity, and specificity of 93.75%, 87.5%, and 100%, respectively. Conclusions: Postoperative cystography features can predict PPI recovery in patients with localized prostate cancer. A deep-learning model can facilitate the identification process. Further validation and exploration are required for the future development of artificial intelligence (AI) in this field.

Keywords: artificial intelligence; cystography; deep learning; postprostatectomy incontinence; radical prostatectomy; urinary incontinence.

Figure 1

Illustration of Cystography feature. (A). Urinary leakage in cystography was defined as contrast leakage at bladder neck and urethra anastomosis site (red arrow). (B). The distance of downward bladder neck was defined as the length of bladder neck below the pelvic inlet lower margin, shown as the length of the red arrow between the two black horizontal lines. (C). The bladder neck angle was measured as the angle of bladder neck to bilateral bladder margin over pelvic inlet. The angle was shown as the angle between the two red lines. (D). The bilateral margin of the ROI images is the bilateral wall of the bladder, as shown in the cystography; the upper margin is the ischial spine and the lower margin is the inferior pubic ramus. Within this defined area, we adjusted the upper and lower margins of the cystography images to fit the horizontal part of the pubic bone in the middle third of the ROI images.

Figure 2

Receiver-operating characteristics curve analysis. (A) ROC curve of downward bladder neck and bladder neck angle for pad-free status 4 weeks after surgery. (B) ROC curves for the best performing deep learning model to predict unfavorable and favorable cystography group.

Figure 3

Nomogram of cystography features to predict pad free probability in 4 weeks.

Figure 4

Gradient-weighted Class Activation Mapping (Grad-CAM). The most important region for making decisions using our model, as determined by the Grad-CAM.