Real-time task recognition in cataract surgery videos using adaptive spatiotemporal polynomials

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Technology

Background:

• Real-time surgical task recognition is crucial for intraoperative guidance.
• Existing methods face challenges with eye motion and zoom variations in surgical videos.

Purpose of the Study:

• To develop and evaluate a new algorithm for real-time surgical task recognition.
• To improve information delivery to surgeons during video-monitored surgeries, specifically cataract surgery.

Main Methods:

• A novel algorithm using adaptive spatiotemporal polynomials for multiscale motion characterization.
• Normalization of videos to compensate for eye motion and zoom.
• Multiple-instance learning for joint adaptation of polynomial basis and key polynomial identification.

Main Results:

• The proposed algorithm achieves improved performance in surgical task recognition (Az = 0.851 vs. 0.794).
• Enhanced accuracy in joint segmentation and recognition of surgical tasks (Az = 0.856 vs. 0.832).
• The system demonstrates real-time processing capabilities.

Conclusions:

• The new algorithm effectively recognizes surgical tasks in real-time, outperforming previous approaches.
• This technology has the potential to enhance surgeon support during complex procedures.
• The method is particularly suitable for deformable objects with fuzzy borders common in surgical videos.