One of the things we are doing differently with Laptitude is using AI technology to generate anatomical models that can be used in our procedural simulations. The AI work has been conducted in-house at Grendel with the special expertise of our very own Luan Duong, working closely with our art Director Anne Draaisma.
Wouter IJgosse, MD PhD of the Radboud University Medical Centre presented the initial evaluation of the generative AI technology to produce realistic training models at the Society of American Gastrointestinal and Endoscopic Surgeons (SAGES) Annual Meeting in Long Beach, California
“The use of generative AI technology to generate novel 3D models for surgical simulation training, has never, to our knowledge been done,” said IJgosse. “We showed generated models of gallbladders and their ducts and arteries to a group of surgeons who rated them as realistic with a great deal of usability in surgical simulation training. I’m very excited to be part of the scientific evaluation of the usability of these new technologies in surgical training and I can see a future where we can use AI to generate the models and scenarios used for simulation training and to measure the performance of trainees in these scenarios. ”
Luan Duong, Grendel –
“There’s lots of work going on looking at how AI can help to create simulation training models based on an individual patient’s CT or MRI scans, but this is the first time that a model has been trained to generate 3D interactable models that incorporate all the available anatomical variation from hundreds of CT or MRI scans. This enables us to create unique, never before encountered combinations of anatomical and pathological variations to test trainee pattern recognition and decision-making skills.”
“Normally a 3D artist needs to model each of these anatomical variations individually and each model can take tens of hours to get right in a realistic and interactable way,” said Anne Draaisma, Art Director, Grendel Medical. “Using the generative AI model in combination with some procedural techniques for essential anatomical details, we can speed the process up enormously and ultimately automate it, so that with each attempt at a simulated procedure, we can generate a unique anatomical puzzle for the trainee to solve ”
Feasibility of AI-Generated Organ Models for Laparoscopic Training: Validation through Expert and Resident Evaluation.
Wouter IJgosse1, MD, PhD, Jan-Maarten Luursema1, PhD; Luan Duong2; Otmar Buyne1, MD, PhD; Harry van Goor1, Em, Prof, Dr; Bas Verhoeven1, MD, PhD.
1Radboud University Medical Centre, Nijmegen The Netherlands 2Grendel Games BV, Leeuwarden, The Netherlands
INTRODUCTION:
Laparoscopic simulation training is essential for surgical trainees before performing laparoscopy in the operating room, where they usually start with laparoscopic cholecystectomy. Although high-fidelity digital simulators are widespread, most fail to implement the anatomical variability and pathologies found in the operating room. This may limitthe transfer of skills to the operating room and delay the development of surgical competence.
To introduce real-world anatomical variations in realistic digital organ models for laparoscopic training, we created cholecystectomy-related anatomical models using artificial intelligence techniques in combination with procedural modeling. This study is the first to establish face validity of artificial intelligence-based digital models.
METHODS AND PROCEDURES:
Large organ models were generated using Latent Diffusion Models trained on liver and gallbladder images segmented from computed tomography scans. Fine structures, including small blood vessels, ductal structures, and organ surface, were enhanced and corrected through an overlaid procedural modeling system, guided by established anatomical variations.
Five different organ sets with varying features were independently evaluated by senior surgical residents and expert gastrointestinal surgeons at the Radboudumc. They scoredtheir agreement with statements pertaining to the morphology, configuration, and relevance to surgical training of the models, on a 5-point Likert scale. Suggestions for improvement were provided in free-text format.
RESULTS:
Twenty-five participants (12 senior residents and 13 expert surgeons) completed the assessment. Twenty-one participants (8 senior residents and 13 experts) had served as the primary or supervising operator in more than 100 laparoscopic procedures. Mean scores for all variables including shape, color, texture, and anatomical positioning were between 3 (neutral) and 4 (good), indicating general acceptance. The configuration of bile ducts and blood supply received the lowest ratings, with a mean score of 3.3. The similarity to the diversity of anatomical variants in the operating room was rated positively, with a mean score of 3.8. Suggestions included enhancing the depiction of disease-specific aspects, such as gallbladder hydrops in cholecystitis. Overall, the models were deemed suitable for integration into surgical skills training, with a mean score of 4.1.
CONCLUSION:
Experienced senior residents and expert surgeons provided favorable ratings for overall morphology, coloration, texture, and anatomical positioning, although the configuration of bile ducts and vascular structures received comparatively lower scores. Participants recognized the models’ ability to represent real-life anatomical variability and deemed them suitable for surgical skills training. These findings suggest that combining artificial intelligence with procedural modeling can generate anatomically varied digital organ models.
Conflicts of Interest:
Wouter IJgosse, MD, PhD and Bas Verhoeven, MD, PhD have received research funding from Grendel Games. Luan Duong is a full-time employee of Grendel Games. The remaining authors declare no other conflicts of interest.
