301 posters, 50 videos, 13 topics, 13 sessions, 734 authors, 193 institutions
ePostersLive by SciGen Technologies S.A. All rights reserved.
ISMICS 2026 Annual Meeting
10 - 13 June, 2026 | Miami, Florida
P53
From Simulation To Surgery: Enhancing Visual Realism In Minimally Invasive Mitral Valve Procedures
Poster Presenter
Authors
Patrick Carnahan, Emma Zhang, John Moore, Terry Peters, Elvis Chen, Gianluigi Bisleri
Affiliations
Cardiac Surgery Department, St. Michael’s Hospital, University of Toronto, Toronto, Canada, Schulich School of Medicine & Dentistry, Western University, London, Canada, Imaging Laboratories, Robarts Research Institute, Western University, London, Canada, Electrical and Computer Engineering, Western University, London, Canada
Part of Topic
Education, Imaging and Facilitating Technologies
Video
Background
●Simulation-based training is increasingly important in competency-based medical education because it allows trainees to build technical skill, judgment, and procedural confidence in a controlled environment before entering the operating room [1].
●Minimally invasive mitral valve repair depends on an endoscopic view, so surgeons must interpret anatomy, depth, tissue appearance, and instrument interactions through the camera rather than direct visualization.
●Silicone mitral valve phantoms are practical for simulation-based training, but their visual appearance differs substantially from the intraoperative field.
●This mismatch may limit trainee immersion and reduce transfer of visual-anatomical skills from the simulator to the operating room.
●Prior work by Engelhardt et al. [2] showed that hyperrealistic image translation can improve mitral valve training phantoms, motivating further work on realism and temporal stability.
Aim
●Develop and validate a surgically relevant image dataset and neural image translation framework to improve the visual realism of mitral valve training phantoms.
●Better replicate the live operative appearance of minimally invasive mitral valve surgery.
●Evaluate both perceptual realism and temporal stability, since useful simulation must support continuous endoscopic viewing rather than isolated still images.
Methods
●Collected endoscopic video data from minimally invasive mitral valve repair procedures and from silicone mitral valve phantoms positioned to match operative viewing angles.
●Curated 59,411 surgical frames from 570 surgical videos and 6,625 phantom frames from the simulator domain.
●Compared unpaired image-to-image translation models including CycleGAN, MUNIT, and DRIT [3].
●Tested patch-based consistency constraints to reduce frame-to-frame flickering during video playback.
●Evaluated perceptual similarity and temporal consistency in both static images and dynamic video sequences
Dataaset and Workflow
●We built a surgical image dataset that captures clinically relevant mitral anatomy, endoscopic lighting conditions, tissue appearance, and instrument interactions seen during minimally invasive repair.
●Phantom images were collected under endoscopic viewpoints chosen to resemble the depth, angle, and field of view seen in the real surgical videos.
●Neural image translation was then applied to convert visually simple phantom images into outputs that more closely resembled the live operative field.
Results
●Image translation improved phantom realism by adding more representative tissue texture, coloration, lighting effects, and instrument context to the simulator.
●Standard style translation produced acceptable realism in static images, but greater visual detail in video sequences was associated with increased frame-to-frame flickering, indicating a trade-off between realism and stability.
●Incorporating temporal consistency constraints reduced flicker and improved output stability while maintaining realism, supporting more continuous and usable endoscopic visualization for training.
●Increasing dataset size was associated with improved image fidelity and better visual continuity across adjacent frames in the translated sequences.
●Image translation improved phantom realism, adding tissue coloration, specular highlights, and more operative-looking visual context.
●Best perceptual realism was achieved by MUNITImgNet, with perceptual similarity (FID) score 16.3 ± 3.2 versus 79.7 ± 17.9 for baseline simulator images.
●CycleGANPatch improved both realism and stability, reducing FID from 42.7 ± 9.1 to 30.2 ± 4.5 while also improving temporal consistency.
●Overall, lower FID scores generally came with greater frame-to-frame variation, highlighting a realism-versus-stability trade-off relevant to video-based simulation.
Conclusions
●AI-based image translation can narrow the visual gap between silicone mitral valve phantoms and the live operative field.
●For clinical simulation, photorealism alone is not enough; stable video output is also required for continuous endoscopic training.
●These findings support integration of image translation into next-generation minimally invasive and robotic mitral simulation platforms.
●Future work will focus on surgeon validation, trainee performance studies, and improved temporally consistent models for real-time deployment.