Evaluating frame selection strategies for video-based organ segmentation algorithms in laparoscopic surgery : a comparison of three strategies

Background: Artificial intelligence (AI)-driven computer vision algorithms can assist in automated anatomical recognition during laparoscopic procedures and enhance both workflow efficiency and intraoperative guidance. Machine Learning (ML) models can be trained to segment anatomical structures, but their performance depends on the quality of annotated data used for training. Frame selection plays a crucial role in optimizing annotation efficiency while maintaining segmentation accuracy. This study compares three frame selection methods – equidistant selection, random selection and active learning (AL) – to determine which strategy yields the best performance for organ segmentation, addressing a gap in the existing literature. Materials & Methods: Two medical students performed pixel-based annotations on videos of exploratory laparoscopies for cancer staging, retrieved from the Ghent University Hospital database. Videos were annotated based on a predefined annotation checklist of anatomical structures and surgical instruments. A pre-trained AL model was updated over three iterative cycles, selecting frames based on uncertainty and diversity measures to optimize annotation efficiency. Segmentation performance was evaluated using three metrics: Intersection over Union (IoU), Precision and Recall. To assess differences between the three frame selection methods, a one-way ANOVA test was performed. Results: Across all anatomical structures and surgical instruments, the mean IoU values were 53.58% for equidistant selection, 52.43% for random selection and 53.2% for AL selection. IoU values for individual anatomical structures and surgical instruments ranged from 37.30% to 65.28%. Statistical analysis showed no significant difference in IoU between the three frame selection methods (p>0.05). Conclusion: The choice of frame selection method did not significantly impact segmentation performance. Simpler methods, such as equidistant and random sampling, achieved comparable results to the more complex AL method. Although AL did not notably enhance performance, it shows potential in reducing annotation effort. Overall, the frame selection strategy does not appear to influence segmentation outcomes for anatomical structures.