(UroToday.com) The second plenary session on the first day of the WCET 2024 Conference was initiated with a moderated panel session by Dr. Andrew J. Hung, MD, on the topic of artificial intelligence (AI) in urologic surgery. This included a comprehensive panel of urologists including Dr. Timothy C. Chang, MD of Stanford Urology, Dr. Ahmed Ghazi, MD, FEBU, MHPE of Johns Hopkins, Dr. Russell Terry Jr., MD of UF Health, and Dr. Bhaskar Somani, MRCS, FEBU, FRCS (Urol) of the University of Southampton.
Following an insightful introduction by Dr. Hung, Dr. Chang began his presentation geared towards “Artificial Intelligence and Endoscopy” with reviewing bladder tumors. For tumor resection, it is widely known that the quality of such resection is dependent upon multiple elements, through crucial to patient prognosis. In an effort to strengthen this quality, various adjunctive technology with white light adaptions in cystoscopy have been evaluated. However, Dr. Chang poses the further development of such diagnostic abilities with AI technology. To this end, a prior colleague of his developed a program entitled “CystoNet”, a deep learning algorithm for papillary bladder tumor detection and classification (Figure 1). With a development set of 100 patients and a validation set of 54 patients, there was a reported performance of 95%.
Figure 1. Tumor detection and classification for CystoNet.
Following this, their team aimed to implement a similar algorithm in real-time mechanisms for intraoperative applications, with primary aims for the AI to classify and segment lesions (ie where normal tissue lies). They completed a benchmark cystoscopy data set that annotated images and videos of both a cancerous and benign region; primarily important for the training, validation, and comparison of AI models, while also serving as an educational atlas for trainees. Further applications of such projects can be incorporated to upper tract imaging and surveillance ureteroscopies. Aside from oncologic practice, Dr. Chang shared his team has translated this program for AI detection of kidney stone composition and patient management.
Dr. Ghazi then proceeded with a presentation on “Artificial Intelligence Applications for Simulation and 3D Modeling”. He focused on two aspects related to performance of the surgeon: predicting (surgical training) and optimizing (operative performance). With simulation training. Dr. Ghazi and his lab have developed near life experiences with a combination of 3D printing, hydrogel molding, and mechanical testing, as well as differentiation between experts and novices with regard to surgical experience. This was proceeded by a project entitled Robust AI-Generated Data-driven Assessment of Robotic Skills (RADAR), a project that included tensile, eye tracking, kinematic, and gesture classification performance data. Furthermore, at the 2022 AUA annual meeting, they utilized 2 XI Robots, where surgeons with an average caseload of 3000 prostatectomies performed an identical procedure. Using machine learning models, they were able to predict surgeon performance and caseload using the model. When combining all data, supervised classification with the combination of surgical console, gesture, and force sensor data achieved high accuracy in determining the completion of literature further defined a nerve sparing learning curve in a physical simulation.
He then went on to deliberate 3D models, in which manual segmentation is quite labor-intensive work, an extensive process from initial imaging to finalized models (Figure 2). With computer vision, they have been able to auto segment CT scans and create intraoperative overlays of such 3D models.
Porpiglia et al. has also been able to accomplish this which resulted in higher enucleation rate, a lower collecting system opening, and fewer postoperative complications. He concluded this portion of the panel by stating, “This is a very instrumental field that is ever expanding, and I can visualize increased applications of this technology”.
The session was continued with Dr. Terry and a presentation entitled “Measurement and Prediction of Kidney Stone Growth Using Machine Learning”. Prior to his discussion, he reviewed the importance of predicting stone growth wherein this can lead to better patient counseling, appropriate use of medical management for prevention, and earlier surgical intervention. As current literature suggests, the ideal metric for stone growth is stone volume. However, stone volume measurements are not yet standardized, although there are different techniques available. Dr. Terry went on to share one of his patient’s stone volume assessments over the course of the year. Although the stone measurements remained consistent on radiologic interpretation, volumetric assessment did show a 60% increase in volume.
Following predicting stone measurement, he transitioned onto predictors of stone growth. He briefly reviewed the components of supervised machine learning, in which an algorithm is trained on a labeled data set with variables. For predicting stone growth, these variables can include CT features, 24-hour urine values, serum lab values, demographics, data from wearable health devices, and any other metrics of interest. At the University of Florida, Dr. Terry and colleagues created a unified model using extracted information from serial CT scans and 24-hour urine data with the ultimate goal of stone growth prediction (Figure 3). At UCSF, they have demonstrated novel machine-learning to predict stone recurrence with 24-hour urine data.
Figure 3. Unified stone growth prediction model created by Dr. Terry and colleagues.
As the final panelist, Dr. Somani shared his talk entitled “Artificial Intelligence in Predicting Outcomes and Management of Kidney Stones”. He initially began by reviewing a study from his group regarding machine learning for predicting stone-free status following shockwave lithotripsy. Here, machine learning algorithms were determined to be adequate, if not more optimal compared to standard approaches. “This means in time they are only going to get better…If this is the case perhaps, we will be using them to make our decisions for shockwave outcomes”. They additionally evaluated a predictive model for post-ureteroscopic urosepsis requiring intensive care admission; overall they found that the model correctly predicted the risk of sepsis in 82% of patients. In another iteration, they reviewed the use of stone volume by machine learning to predict stone free status at ureteroscopy. With 330 patients, variables of importance included stone volume (17.7%), operative time (14.3%), age (12.9%), and stone composition (10.9%). Equipped with tools to predict stone free status, surgeons are able to adequately counsel patients preoperatively. He concluded his presentation by stating, “I think there is a big future for AI in imaging, stone composition, lithotripsy, PCNL, spontaneous stone passage, ureteroscopy, and predicting outcomes of procedures.
Dr. Hung then concluded the session with the following statement: “Clearly, there is tremendous work that is rigorous using AI not just as a trendy pocket but geared towards patient care. I encourage our community, the one that has adopted so many technologies, to seriously identify how AI can help with our patient care.”
Presented by:
Moderator: Timothy C. Chang, MD – Artificial Intelligence and Endoscopy
Expert Panel:
- Timothy C. Chang, MD – Artificial Intelligence and Endoscopy
- Ahmed Ghazi, MD, FEBU, MHPE - Machine Learning and 3D Models for Predicting Performance of Robotic Surgery
- Russell S. Terry, Jr., MD – Measurement and Prediction of Kidney Stone Growth Using Machine Learning
- Bhaskar Somani, MRCS, FEBU, FRCS(Urol) – Artificial Intelligence in Predicting Outcome and Management of Kidney Stones
Written by: Mariah Hernandez, Research Specialist, Department of Urology, University of California Irvine, @mariahch00 on Twitter during the 2024 World Congress of Endourology and Uro-Technology: August 12 -16, 2024, Seoul, South Korea