Residual Motion Improves Deep Learning Performance for Surgical Actions in Gynecologic Laparoscopy

In a recent work, presented at the 31st IEEE International Symposium on Computer-Based Medical Systems (CBMS2018), we could show that the inclusion of Residual Motion improves classification performance of surgical actions in videos from gynecologic laparoscopy significantly (resulting in a boost of Recall and Precision of 5% and 9% with the GoogLeNet CNN architecture). This performance can be improved even further (to a boost of 13% and 25% in terms of Recall and Precision) by using a late fusion approach for frame classification in the videos. The corresponding paper can be found here.

Comments are closed.