Improving situation recognition using endoscopic videos and navigation information for endoscopic sinus surgery.

Area of Science:

• Medical technology
• Surgical robotics
• Artificial intelligence in medicine

Background:

• Endoscopic sinus surgery (ESS) requires high surgeon skill due to the narrow surgical field and proximity to vital structures.
• Previous work developed an automatic situation recognition method for ESS using endoscopic images and navigation data.

Purpose of the Study:

• To enhance the accuracy of automatic surgical situation recognition in ESS.
• To improve upon a previously developed recognition method by refining model parameters and incorporating post-processing techniques.

Main Methods:

• Optimized training model parameters and Long Short-Term Memory (LSTM) units.
• Modified data augmentation techniques and implemented a one-dimensional median filter for post-processing.
• Evaluated the enhanced method using clinical data from ESS procedures.

Main Results:

• Overall scene recognition accuracy was improved compared to the previous study.
• A one-dimensional median filter significantly reduced short-time false recognitions.
• Post-processing with scene transition constraints was necessary for further accuracy gains, though phase recognition saw limited improvement.

Conclusions:

• Model parameter tuning, 1D filtering, and post-processing can enhance ESS situation recognition.
• Current accuracy levels are still unsatisfactory, indicating a need for more advanced recognition and post-processing methods.
• Further research is essential to achieve highly accurate and reliable automatic situation recognition in ESS.