A Visualization Training Tool for Optimum Cutting Range During Urethral Dissection in Robot-Assisted Radical Prostatectomy: A deep learning approach

Abstract

Surgical training and skills assessment is one of the methodologies used by clinical practitioners to master their skills in a fault-tolerant, safe and risk-free environment. Software training tools is one of them. Narrowing the scope on surgery, and with the significant popularity of robotic surgeries in the last few years, clinical and technological research has shifted their attention towards developing surgical supportive tools such intraoperative surgical planning, training tools, etc. In this thesis, we attempt to develop a software to teach surgeons on the optimum urethra dissection locations based on prostate size and cancer location. The surgeon is presented by a patient case and suggests optimum different dissection locations. The system would automatically evaluate the surgeon's performance and generate a score accordingly. In this thesis, a clinical validation was conducted to validate the need of such software in a clinical practice by interviewing three robotic surgeons experts. Once validated successfully, Tutorial Module was developed and validated with expert urology surgeons using content validity to evaluate the module's effectiveness as a teaching modality and Questionnaire for User Interaction Satisfaction to evaluate the interface perceptiveness among them. Both tests reflected high surgeons' satisfaction. To detect the urethra, U-Net, a deep learning module was trained over a small dataset of images extracted from from HMC's operating rooms as well as publicly available videos. Image semantic segmentation was the method used to extract and locate the urethra. U-Net model was able to reach a Jaccard Coefficient of 78.42% and a Dice Coefficient of 88.11%. To the best of our knowledge, this thesis is the first of its kind to offer a scenario-based training on optimum urethra dissection location for robot-assisted radical proastatectomy