Machine learning has gained significant popularity in recent years, but this technology is not new to urology. In 1997, Kattan and colleagues explored the use of artificial neural networks for prostate cancer survival prediction.¹ Since then, disruptive technologies and advancements in computational processing power have enabled the expansion of machine learning techniques into previously unattainable realms. Contemporary applications of artificial intelligence (AI) tools are widespread and include prostate cancer imaging, pathology interpretation, and even robotic surgery.

Figure. Summary of surgical gestures during performance of vesicourethral anastomosis in robotic-assisted radical prostatectomy; gestures differentiate novice from expert surgeons and are associated with greater suturing efficiency and lower tissue trauma. Reprinted with permission from Chen et al, J Urol. 2018;200(4):895-902.⁵

Among the ripest targets for the use of AI in prostate cancer is to aid in prostate cancer diagnosis on MRI. Prostate anatomy is often segmented as part of radiologists’ routine clinical workflow to enable MR-fusion prostate biopsy. This practice has facilitated the rapid development of large and high-quality data sets with minimal manual data labelling, which can often be labor intensive and in some cases prohibitive. Further, correlative pathologic data are often available from both prostate biopsy as well as subsequent prostatectomy, providing AI models with a clear and robust “ground truth” upon which to train. Several groups have explored the use of AI to interpret prostate MRI imaging with quite promising results.² Similarly, AI has shown promise in interpretation of prostate biopsy histopathology. Paige Prostate, a commercial AI software for automated interpretation of prostate biopsy pathology, was the first AI-powered pathology interpretation algorithm to receive Food and Drug Administration approval.³

In recent years, applications of AI in prostate cancer have moved beyond imaging and pathology, and into surgical interventions including robotic-assisted radical prostatectomy (RARP). In their 2018 study, Hung and colleagues demonstrated that automated technical performance metrics of surgeons were correlated to hospital length of stay following RARP.⁴ This landmark study provided proof-of-concept that intraoperative events were empirically evaluable and meaningfully associated with patient outcomes. Hung et al subsequently demonstrated a correlation between surgical performance metrics with surgeon skill level (see Figure),⁵ recovery of urinary continence,⁶ and erectile function⁷ following RARP. Similarly, Schuler and colleagues demonstrated that a series of performance metrics including surgical gestures, robotic instrument kinematics, and tissue force were predictive of individual surgeon expertise in RARP.⁸

AI also has promising applications in robotic surgery beyond surgical skills assessment. Our group recently developed a novel AI-powered computer vision platform for fully automated detection of key surgical steps in RARP.⁹ This AI tool is capable of accurately identifying sequential steps of RARP, such as space of Retzius dissection, anterior and posterior bladder neck dissection, seminal vesicle/posterior dissection, vesicourethral anastomosis, etc. This work is distinct from prior efforts to apply AI to robotic surgery in 2 key areas: (1) AI step detection is based purely on video footage alone without inputs from the surgical platform or instruments, and (2) our AI model moves beyond microgestures and instead assesses entire phases and steps of surgery, taking global anatomic and temporospatial relationships into consideration to provide meaningful predictions of surgical phase.

The potential applications of a robust AI step detection tool for RARP are myriad. Not only does comprehensive step detection lay the foundation for future efforts to continue correlating intraoperative events with postoperative outcomes, but step detection also serves as the engine to drive innovative AI applications in surgical training and education, quality and safety benchmarking, medical documentation, and operating room logistics. As a proof-of-concept using our RARP step detection algorithm, we recently developed a novel AI-based tool for generating operative reports for RARP based purely on full-length surgical video footage alone.¹⁰ Notably, AI-generated operative reports in RARP achieved similar accuracy to actual operative reports written by surgeons, thus demonstrating the feasibility of AI-driven technology in robotic surgery to potentially improve surgical workflows, reduce documentation burden, and enhance report accuracy.

Building upon the numerous exciting advancements in AI for prostate cancer over the last several years, the future holds tremendous potential for transformative innovations in robotic surgery. This includes real-time surgeon feedback and intraoperative decision support, which have the potential to revolutionize the experience of robotic surgery for surgeons and drive improvements in outcomes for patients. However, as with all new technologies, AI in prostate cancer care must be developed and implemented with poise and balance. Specific challenges that our field must address include bias inherent within training data sets, explainability of “black-box” AI models, external validity to diverse practice settings, and always maintaining a “human in-the-loop” to prevent erosion of the surgeon-patient relationship that is fundamental to the practice of medicine.