Related Experiment Video
Updated: Jan 6, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Weakly supervised pre-training for surgical step recognition using unannotated and heterogeneously labeled videos
Sreeram Kamabattula1, Kai Chen1, Kiran Bhattacharyya2
1Advanced Product Development, Intuitive Surgical, Inc., 5655 Spalding Drive, Peachtree Corners, GA, 30092, USA.
International Journal of Computer Assisted Radiology and Surgery
|December 2, 2025
Summary
Weakly supervised pre-training using surgical phase labels significantly enhances automated surgical step recognition in minimally invasive surgery training, even with limited annotations. This method improves scalability for surgical education and feedback systems.
Area of Science:
- Computer Vision
- Medical Education Technology
- Surgical Robotics
Background:
- Manual annotation of surgical videos is crucial for training but is time-consuming and limits scalability.
- Automated surgical step recognition is vital for providing feedback and assessment in surgical training.
- Current methods struggle with limited annotated data in real-world surgical training scenarios.
Purpose of the Study:
- To propose and evaluate a weakly supervised pre-training framework to improve automated surgical step recognition.
- To leverage unannotated or heterogeneously labeled surgical videos for better model performance.
- To address the challenge of data scarcity in surgical video annotation for training.
Main Methods:
- Evaluated three types of weak labels: surgical phases, cross-procedure steps, and intraoperative time progression.
- Utilized datasets from four robotic-assisted procedures (sleeve gastrectomy, hysterectomy, cholecystectomy, radical prostatectomy).
- Simulated annotation scarcity by varying the proportion of available step annotations (α ∈ 0.25, 0.5, 0.75, 1.0) and benchmarked a 2D CNN model with/without weak label pre-training.
Main Results:
- Pre-training with surgical phase labels (PHASE-WITHIN) consistently improved step recognition, with up to 6.4 f1-score points gain under limited annotations (α=0.25).
- Cross-procedure step pre-training and time-based labels offered benefits depending on the procedure.
- Label efficiency analysis indicated that weak pre-training matched performance requiring significantly more manual annotations.
Conclusions:
- Weakly supervised pre-training is a practical strategy to enhance surgical step recognition with scarce annotated data.
- This approach supports scalable feedback and assessment in surgical training.
- It offers a feasible solution for surgical training workflows where comprehensive annotations are impractical.
