Related Experiment Video
Updated: Feb 20, 2026
03:27
The Role of Indocyanine Green Fluorescence in Complex Laparoscopic Cholecystectomy Navigation
Published on: January 31, 2025
1.4K
CVS assessment via distillation-based self-supervised and multiple instance learning in laparoscopic cholecystectomy
Hao Wang1,2,3, Yutao Zhang4,5, Yuxuan Yang4,5
1School of Management, Hefei University of Technology, 193 Tunxi Road, Hefei, 230009, Anhui, China. haowang@hfut.edu.cn.
International Journal of Computer Assisted Radiology and Surgery
|February 18, 2026
Summary
This study introduces the SMIL framework for automated critical view of safety (CVS) assessment during laparoscopic cholecystectomy (LC). SMIL enhances surgical safety by accurately identifying CVS without needing costly segmentation labels.
Area of Science:
- Medical Imaging and Artificial Intelligence
- Surgical Safety and Analytics
Background:
- Accurate assessment of the critical view of safety (CVS) is vital for preventing bile duct injuries during laparoscopic cholecystectomy (LC).
- Current automated methods often require expensive segmentation annotations and struggle with temporal-spatial understanding, limiting their real-world application.
- There is a need for efficient and robust frameworks for automated CVS assessment that overcome these limitations.
Purpose of the Study:
- To develop an efficient framework for automated CVS assessment in LC that eliminates the need for segmentation annotations.
- To enhance model robustness and improve spatiotemporal comprehension for better surgical safety analysis.
- To establish a new benchmark for label-free automated CVS assessment.
Main Methods:
- Introduction of the SMIL framework, combining distillation-based self-supervised pretraining and multiple instance learning (MIL).
- A video transformer was pretrained using label-free self-distillation to extract spatiotemporal features.
- The framework was fine-tuned using MIL, integrating global and local representations for multi-label CVS classification on the Endoscapes2023 dataset.
Main Results:
- The SMIL framework demonstrated superior performance compared to state-of-the-art methods on the Endoscapes2023 dataset without segmentation labels.
- SMIL achieved significant improvements over the strongest label-free baseline, with gains of 3.21% in mean average precision (mAP) and 2.74% in balanced accuracy.
- Notably, SMIL surpassed even segmentation-supervised models in mAP, highlighting its efficient learning capabilities.
Conclusions:
- The SMIL framework provides an effective solution for automated CVS assessment without requiring segmentation annotations or sequential inputs.
- By integrating self-supervised learning and MIL, SMIL enhances spatiotemporal understanding and generalization in the context of LC surgeries.
- This framework offers significant theoretical insights and practical value for improving surgical safety and preventing bile duct injuries.
