SurgflowNet: Leveraging unannotated video for consistent endoscopic pituitary surgery workflow recognition

Area of Science:

• Neurosurgery
• Computer Vision
• Artificial Intelligence

Background:

• Automating surgical workflow recognition is crucial for training and efficiency.
• Endoscopic pituitary surgery presents unique challenges due to complex, variable workflows.
• Current methods struggle with data scarcity and transferability in pituitary surgery.

Purpose of the Study:

• To develop a novel model for consistent endoscopic pituitary surgery workflow recognition.
• To leverage unannotated data to overcome annotation limitations.
• To improve the accuracy and robustness of surgical workflow recognition in complex procedures.

Main Methods:

• Developed SurgflowNet, a spatio-temporal model for workflow recognition.
• Utilized a limited annotated dataset to infer quasi-labels for unannotated videos.
• Employed a student-teacher framework for frame encoder training and a consistency loss for stability.

Main Results:

• SurgflowNet demonstrated significant improvements over state-of-the-art methods.
• Achieved a 5% increase in macro F1-score and a 13.4% increase in Edit Score.
• Successfully leveraged unannotated data for robust model training.

Conclusions:

• SurgflowNet offers a robust solution for endoscopic pituitary surgery workflow recognition.
• The model's ability to use unannotated data addresses a key challenge in the field.
• This advancement has the potential to enhance surgical training and intraoperative efficiency.