Automated multimodel segmentation and tracking for AR-guided open liver surgery using scene-aware self-prompting.

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Augmented Reality (AR) integration in surgery requires real-time, accurate visual data processing.
• Open liver surgery presents complex anatomical variations demanding robust segmentation and tracking.

Purpose of the Study:

• To develop a multimodel, real-time semantic segmentation and tracking system for AR-guided open liver surgery.
• To leverage foundation models and scene-aware re-prompting for balancing accuracy and speed in surgical AR applications.

Main Methods:

• Integrated ESANet (RGBD model), SAM (segmentation foundation model), and DeAOT (video object segmentation).
• Developed an auto-promptable pipeline with a scene-aware re-prompting algorithm adapting to surgical scene changes.
• Evaluated on intraoperative RGBD videos from 10 open liver surgeries using a head-mounted AR device.

Main Results:

• The multimodel approach achieved 71% median IoU at 13.2 FPS without re-prompting.
• Outperformed individual models, offering superior segmentation accuracy over ESANet and better temporal resolution than SAM.
• Scene-aware re-prompting reached 74.7% IoU at 11.5 FPS, matching DeAOT's performance.

Conclusions:

• The scene-aware re-prompting strategy effectively balances segmentation accuracy and temporal resolution for real-time AR liver surgery.
• Integrating complementary models ensures robust and accurate segmentation in complex surgical environments.