Live laparoscopic video retrieval with compressed uncertainty.

Area of Science:

• Medical informatics
• Computer vision
• Surgical technology

Background:

• Keyword-based retrieval is limited for complex medical data like videos.
• Content-based retrieval using media as queries offers a solution.
• Surgical video retrieval is valuable for clinical applications, especially in real-time.

Purpose of the Study:

• To develop a robust and efficient surgical video-to-video retrieval method.
• To address the challenge of unreliable hash bits in video hashing due to temporal fluctuations.
• To improve the accuracy and generalizability of real-time surgical video search.

Main Methods:

• Implemented a novel video hashing technique to convert surgical videos into compact binary arrays.
• Developed a method to mitigate uncertainty in hash bits while maintaining computational efficiency.
• Utilized a multi-task evaluation protocol across six surgery types, including cholecystectomy and bypass phases.

Main Results:

• Achieved superior retrieval results with a 3%-4% improvement in top 10 mean average precision.
• Demonstrated the method's effectiveness on a new, diverse surgical video dataset.
• Validated the generalizability of the approach across various surgical procedures.

Conclusions:

• The proposed method significantly enhances surgical video retrieval accuracy.
• The approach is computationally efficient and generalizable across different surgical contexts.
• This advancement holds potential for real-time decision support in operating rooms.