Query-by-example surgical activity detection.

Area of Science:

• Surgical robotics and automation
• Machine learning in healthcare
• Biomedical signal processing

Background:

• Automating surgical skill evaluation and teaching is crucial for advancing medical training.
• Current methods for analyzing surgical tool motion data are hindered by extensive manual pre-processing.
• Efficiently searching large surgical datasets for specific activity segments is a significant challenge.

Purpose of the Study:

• To develop a content-based information retrieval method for automatic detection of surgical activity segments.
• To enable scalable analysis of long-duration surgical data without manual pre-processing.
• To introduce a query-by-example (QBE) approach for identifying relevant surgical maneuvers.

Main Methods:

• Utilized unsupervised feature learning with a stacked denoising autoencoder (SDAE).
• Employed asymmetric subsequence dynamic time warping (AS-DTW) and template matching for scoring candidate segments.
• Evaluated performance using recall, precision, F1-score, and Jaccard index on JIGSAWS and MISTIC-SL datasets with a suture throw maneuver as the query.

Main Results:

• Achieved high recall (up to 93%) and precision (up to 93%) on the JIGSAWS dataset across different experimental setups (SSST, SSDT, DS).
• Demonstrated robust performance on the MISTIC-SL dataset, with recall up to 87% and precision up to 72%.
• The QBE method successfully identified activity segments matching the query maneuver.

Conclusions:

• A novel content-based information retrieval method was developed for automated detection of surgical activities.
• The QBE approach effectively identifies multiple instances of specific activities within extensive surgical recordings.
• The method shows promise for scalable and automated analysis in surgical training and skill assessment.