Accurate Detection of Out of Body Segments In Surgical Videos using - PowerPoint PPT Presentation

Accurate Detection of Out of Body Segments In Surgical Videos using Semi-Supervised Learning Maya Zohar, Omri Bar, Daniel Neimark, Gregory D. Hager, Dotan Asselmann theator Inc., San Mateo, CA, USA Department of Computer Science, Johns Hopkins University, Baltimore, USA.

Introduction Laparoscopic surgery and Surgical Intelligence

Motivation • Anonymization • Reduce storage size • Enhance model performance

Motivation Irrelevant frames and out of body segments Blurred Dark Out-of-body

Our goal Train a model capable of accurately detecting irrelevant segments throughout an entire surgical video

Quite simple task for a supervised classi fi cation… But what if the data is only partially labeled?

Method Annotate Train Iteration 0 Repeat

Method Annotate Train Iteration 1 Repeat

Method Annotate Train Iteration 2 Repeat

Dataset 640 videos Train Valid Test 600 20 20 From 6 di ff erent Videos Videos Videos medical centers Fully Fully Partially annotated annotated annotated

Results Out-of-body accuracy - 99.85%

Results Out-of-body - 98.8% In-body - 96.5% 97% Recall @ Out-of-body - 97.5% In-body - 96.2% 83.5% Precision

Results Theator, 2020 Twinanda et al., 2014 * 97% Recall @ 56.4% Recall @ 30.5% Precision 83.5% Precision * Andru Putra Twinanda, Michel De Mathelin, and Nicolas Padoy. Fisher kernel based task boundary retrieval in laparoscopic database with single video query. In International conference on medical image computing and computer-assisted intervention, pages 409– 416. Springer, 2014b.

Results Accurate Predictions

Results Examples of Misclassi fi cation

Conclusions Highly accurate classi fi cation of Limitations - handling edge out of body frames cases

Thank you! Gregory Hager Dotan Asselmann Omri Bar Maya Zohar Daniel Neimark maya@theator.io