Abstract
The work presented in this thesis tackles the problem of video analysis for laparoscopic interventions, in the case of very scarcely annotated datasets. During laparoscopic surgery, the live video feed from inside the abdominal cavity of the patient is the keystone of the entire procedure, providing all the visual feedback required by the surgeon and the staff to carry out the intervention. Our objective is to develop methods capable of leveraging this source of information and understand it; thus providing the foundation for context-aware, vision-based systems to be developed in the future for decision support in the operating room.
Recent breakthroughs in computer vision, spearheaded by deep learning methods brought major advances to surgical video analysis. However the current default approach of full supervision is an obstacle for future developments. Laparoscopic procedures generate vast amounts of video data which, even if stored, will for the most part almost certainly remain unannotated. This thesis investigates the use of unannotated data across multiple visual tasks. In our first approach, we propose an automatic annotation method relying on a very small ratio of manually annotated data, and demonstrate the usability of the automatically annotated data for training real-time CNN-LSTM predictors. Our second approach then shows how this unannotated data can be leveraged and explored in a scalable, OR-compatible manner and without any annotations using video hashing. By learning in a self-supervised manner, searchable binary representations of surgical videos, we are able to retrieve video content matching a given scene, represented by a video clip, in terms of surgical phase or surgical critical event.
Jury members
- Prof. Nicolas Padoy, University of Strasbourg, IHU Strasbourg, France
- Prof. Didier Mutter, University Hospital of Strasbourg, France
- Dr Marie-Odile Berger, Research Director, INRIA Nancy
- Prof. Danail Stoyanov, Professor, University College of London
- Dr Pierre Jannin, Research Director, INSERM Rennes