Related Experiment Video
Updated: Mar 7, 2026
Retzius-Sparing Robot-Assisted Radical Prostatectomy
Published on: May 19, 2022
Human-in-the-loop validation of an artificial intelligence-driven real-time prostatic capsule recognition model for
Wei Chen1,2, Shohei Fukuda3, Soichiro Yoshida1
1Department of Urology, Institute of Science Tokyo, 1-5-45 Yushima, Bunkyo-Ku, Tokyo, 113-8519, Japan.
Background:
Nerve-sparing robot-assisted radical prostatectomy (NS-RARP) requires precise prostatic capsule identification to balance oncological control with functional preservation. Artificial intelligence (AI)-based intraoperative navigation could provide real-time anatomical guidance, but clinical feasibility remains unvalidated.
Methods:
We conducted a human-in-the-loop feasibility study evaluating a convolutional neural network system for real-time fluorescence-like navigation (FLN) prostatic capsule identification. A total of 285 frames from 12 NS-RARP procedures have been used to establish the AI model. The validation and clinical acceptance assessment were performed on 23 representative frames from 3 NS-RARP procedures by four surgical scenarios: initial lateral dissection (Phase I), visual interference from bleeding/smoke (Phase II), limited capsule exposure (Phase III), and distant field view (Phase IV). 6 experienced urologic surgeons independently evaluated AI output using structured 10-point scales across 6 clinical domains. Quantitative assessment employed standard segmentation metrics comparing AI output to surgeon-annotated ground truth.
Results:
The AI system achieved consistently high clinical acceptance (surgeon ratings 8.21-9.01/10) despite moderate segmentation performance (Dice 0.533 ± 0.098). Boundary positioning accuracy received the highest rating and significantly exceeding other domains (8.63 ± 0.61). Performance varied across surgical scenarios with mean Dice scores of 0.622 in Phase I, 0.489 in Phase II, 0.616 in Phase III, and 0.390 in Phase IV, yet surgeon reliability ratings remained high across all phases (8.39-9.10/10). Weak positive correlation between Dice scores and clinical ratings was observed (r = 0.42, p = 0.048). The system demonstrated high recall of 79.7% with moderate precision of 40.9%, a pattern preferred by surgeons for safety.
Conclusions:
AI-driven real-time prostatic capsule navigation demonstrates clinical feasibility for pilot trials despite moderate technical performance. Surgeons exhibited well-calibrated trust across varying conditions, prioritizing anatomically correct positioning over pixel-perfect segmentation.
