PhD defense: Luca SESTINI
Title: Methods for Learning Surgical Instrument Segmentation from Unlabelled Datasets Using Prior Knowledge
Team: RDH
Abstract: Surgical endoscopic videos represent a rich source of information for analyzing minimally invasive surgical procedures. Identifying and localising surgical instruments from these videos is a crucial step for the development of valuable applications like automatic surgical skill assessment and real-time decision support, aimed at improving the quality of surgical care. Most of the available solutions tackling this problem use fully-supervised learning approaches to train deep learning models on manually annotated data. Due to the cost of manual annotations, the training of such models is confined to limited sets of labelled and curated data, potentially impacting their generalization ability to perform on real-world data. This thesis proposes methods for instrument localisation and identification which can be trained on unlabelled datasets. General knowledge about surgical instruments - cheaper and more easily obtained than manual annotations - is incorporated into the training architectures to fabricate effective supervision signals. The proposed approaches leverage novel methods for unsupervised learning, self-supervised representation learning, and learning from noisy labels, all designed to effectively mine such prior and complementary knowledge. We hope our proposed approaches can facilitate the development of valuable assistive technologies to improve the quality of surgical care.
The presentation will take place on June 8, 2023, at 1 PM, and will be held in Salle Panacée (IHU, 1 Pl. de l'Hôpital, 67000 Strasbourg).
Thesis commitee:
Le dépôt des candidatures pour les postes d’enseignants-chercheur est ouvert. Les offres sont...