Fisher kernel based task boundary retrieval in laparoscopic database with single video query.

Area of Science:

• Medical Informatics
• Computer Vision
• Surgical Technology

Background:

• The increasing volume of laparoscopic videos necessitates efficient search capabilities for training and quality assessment.
• Current video search engines often return entire videos, not specific surgical task segments.
• Manual annotation of large video datasets is time-consuming and impractical.

Purpose of the Study:

• To develop a coarse-to-fine video indexing approach for precise task localization in laparoscopic surgery videos.
• To evaluate the effectiveness of Fisher kernel (FK) encoding for video similarity measurement in this context.

Main Methods:

• A coarse-to-fine video indexing method was proposed to identify time boundaries of surgical tasks.
• The approach utilizes video snippet queries for task identification.
• Fisher kernel (FK) encoding was combined with similarity measures, outperforming traditional methods like dynamic time warping (DTW).

Main Results:

• The method successfully identified 3 tasks in 49 bypass surgery videos.
• It located 1 task in 23 hernia surgery videos.
• A cross-surgery task was identified between bypass and sleeve gastrectomy videos, demonstrating robustness.

Conclusions:

• The proposed video indexing approach effectively localizes specific surgical tasks within laparoscopic videos.
• Fisher kernel encoding offers a superior similarity measure for laparoscopic video analysis compared to DTW.
• This technology has significant potential for enhancing surgical education and video-based quality improvement.