Large-scale surgical workflow segmentation for laparoscopic sacrocolpopexy.

Area of Science:

• Minimally invasive surgery
• Surgical robotics and automation
• Medical image analysis

Background:

• Laparoscopic sacrocolpopexy is the gold standard for vaginal vault prolapse.
• Automated surgical workflow analysis is crucial for studying surgical skills and optimizing procedures.
• Existing benchmarks like Cholec80 do not adequately represent the unique challenges of sacrocolpopexy, particularly its longer and more variable phase durations.

Purpose of the Study:

• To investigate automated surgical workflow segmentation for laparoscopic sacrocolpopexy.
• To address the challenges posed by highly variable and extended phase durations in this procedure.
• To evaluate the suitability of sequence-to-sequence (seq2seq) models for sacrocolpopexy workflow analysis.

Main Methods:

• Introduction of sequence-to-sequence (seq2seq) models for coarse-level phase segmentation.
• Testing multiple architectures including LSTM and Transformer models.
• Utilizing different configurations (time-shifted, time-synchronous) and training strategies.

Main Results:

• 7-fold cross-validation performed on 14 sacrocolpopexy videos.
• Frame-based (accuracy, F1-score) and event-based (Ward metric) evaluations conducted.
• Different architectures showed trade-offs: LSTM/Mode average for frame accuracy, Transformer for phase transition consistency.
• Performance on Sacrocolpopexy differed from the Cholec80 dataset.

Conclusions:

• Workflow segmentation of Sacrocolpopexy presents unique challenges compared to Cholec80, requiring dedicated approaches for longer phase durations.
• Seq2seq models are feasible for Sacrocolpopexy, offering a flexible framework for further exploration.
• Event-based evaluation metrics provide valuable insights complementary to traditional metrics like accuracy and F1-score for workflow segmentation.