Related Experiment Video
Updated: Oct 23, 2025
07:27
Robotics in Surgery: A Modular Robotic Platform Driven Gastric Wedge Resection
Published on: February 7, 2025
691
Surgical workflow recognition with 3DCNN for Sleeve Gastrectomy
Bokai Zhang1, Amer Ghanem2, Alexander Simes2
1C-SATS, Inc. Johnson & Johnson, 1100 Olive Way, Suite 1100, Seattle, WA, 98101, USA. bzhang29@its.jnj.com.
International Journal of Computer Assisted Radiology and Surgery
|August 20, 2021
Summary
This study introduces a deep 3D convolutional neural network (3DCNN) for surgical workflow recognition, outperforming traditional methods. The approach uses focal loss and prior knowledge filtering to improve accuracy and provide smooth predictions in surgical videos.
Area of Science:
- Computer-assisted surgery
- Medical image analysis
- Surgical workflow recognition
Background:
- Surgical workflow recognition is vital for computer-assisted surgery systems.
- Current methods often use Convolutional Neural Network-Recurrent Neural Network (CNN-RNN) architectures.
- These methods face challenges with complex surgical data and imbalanced datasets.
Purpose of the Study:
- To explore the use of deep 3D Convolutional Neural Networks (3DCNN) for surgical workflow recognition.
- To address challenges in surgical workflow recognition, including imbalanced data.
Main Methods:
- Implementation of an Inflated 3D ConvNet (I3D) architecture for surgical workflow recognition.
- Utilization of focal loss (FL) to mitigate imbalanced data issues.
- Application of prior knowledge filtering (PKF) to refine recognition results.
Main Results:
- The proposed I3D-FL-PKF workflow achieved 84.16% frame-level accuracy on a sleeve gastrectomy dataset.
- A weighted Jaccard score of 0.7327 was reached, outperforming CNN-RNN designs.
- Focal loss effectively addressed data imbalance, and PKF smoothed predictions, enhancing overall accuracy.
Conclusions:
- The deep 3DCNN workflow provides consistent and smooth predictions for surgical phases and transitions.
- The integration of focal loss and prior knowledge filtering demonstrates significant potential for clinical applications.
- This approach offers a promising solution for accurate surgical workflow recognition.
