Semiautomated Method for Editing Surgical Videos

Introduction: Surgical videos are increasingly utilized for trainee education, publications, and conferences. Video editing remains a largely manual, laborious, and time intensive process. In this study, we offer a semiautomated method for video editing intended to enhance efficiency and reduce workload. The objective was to create an edited video that includes key scenes to accurately summarize the surgery.

Methods: Five full-length transsphenoidal endoscopic pituitary surgeries were included. These full-length videos were manually edited by surgical residents and served as the gold standard. Once the manual videos were completed, the full-length videos were run through Magisto (www.magisto.com), an artificial intelligence home video editing software. A novel postproduction video editing algorithm developed by our group was then used to remove noninformative scenes from Magisto videos (i.e., endoscope obscured by secretions; [Fig. 1]). This software utilizes a self-supervised K-means classification method to further optimize video segmentation. Magisto and postproduction edited videos were evaluated with a confusion matrix for informative and non-informative scenes. Videos were also qualitatively assessed by three surgeons with a 5-point Likert's scale.

Results: Full-length videos had a mean length of 1 hour and 53 minutes. Magisto edited videos had a mean length of 8 minutes and 26 seconds and postproduction software further reduced videos to a mean length of 4 minutes and 12 seconds, cutting original runtime duration by 93 and 96%, respectively. The manually edited videos had 56 informative scenes and 17 noninformative scenes. Magisto included 32.5 informative scenes (sensitivity = 58.0%) but also 574 noninformative scenes (71.4% of which were endoscope lens irrigation scenes) ([Table 1A]). The postproduction program correctly identified 1,321 out of 1,466 noninformative frames (specificity = 90.1%) to remove from Magisto videos and correctly kept 4,624 out of 4,876 informative frames (sensitivity = 94.8%; [Table 1B]). Experts agreed (Likert's score = 4) with the statement “the overall quality of the video is adequate to share with peers in its current state” 60% of the time. The remaining 40% of the responses were split evenly between neutral ratings (Likert's score = 3) and disagreement (Likert's score = 2) with that statement.

Conclusion: Magisto automatically edited surgical videos captured more than half of the important scenes, but it included a disruptive number of noninformative scenes. These results demonstrate the utility of such software, but also highlights the potential for improvement. Our novel editing algorithm using a self-supervised K-means classification method improved the specificity of Magisto videos. Future refinement of our software with a goal to improve on Magisto will offer a more sensitive and specific, fully automated video editing software geared for surgical videos.

Fig. 1 Diagram of methods.

Table 1 (A) Confusion matrix comparing Magisto with gold standard of manually edited videos. (B) Confusion matrix comparing novel postproduction editing algorithm with manually determined scene importance.

 

MissingLinkText

No conflict of interest has been declared by the author(s).

Publication History

Article published online:
12 February 2021