Semi-supervised deep learning for small annotated databases. Application to the segmentation of multi-modal medical images
Summary: The segmentation of tumor volumes and healthy organs on medical images is an essential step in radiotherapy to maximize the radiation of cancers and at the same time protect healthy tissues and organs. In clinical routine this segmentation is done manually by doctors with a mouse on a computer screen, is time-consuming and can vary greatly from one expert to another. Automatic segmentation is therefore of major clinical interest. In this doctoral thesis, we will focus on the segmentation of medical images, which allows to segment the various organs, tumors or areas inflamed or necrotic in 3D.
The objective of this PhD thesis is to study deep learning methods in the case where we have little data or unreliable annotated data to segment medical images. Deep learning is currently a very active field of research, particularly in medical imaging. In the case of supervised learning, it consists in learning characteristics from an annotated database, ie containing for each observed data, the decision taken by a human intervener. In the case of semi-supervised learning, we have both unannotated data and annotated data. In addition, the number of annotated data may be small compared to non-annotated data and the annotations of the data may be unreliable. Our strategy is to use non-annotated data for improving the reliability of annotations.
The PhD student will work with Quantif team of the LITIS laboratory at the University of Rouen. The Quantif team is multidisciplinary. It collaborates closely with the Henri Becquerel Center which is medical center for fighting against cancer.
Profil: We are looking for an excellent student, with a Master's degree in Applied Mathematics, signal and image processing or equivalent field. Obtaining a Master's degree with a good ranking within the promotion is required. This one must have very good skills in numerical analysis, inferential statistics and computer programming (Python, C ++ or Java). Good knowledge of scientific English is essential.
Contacts: Mcf. Jérôme Lapuyade-Lahorgue: email@example.com et Pr. Su Ruan: firstname.lastname@example.org
(c) GdR 720 ISIS - CNRS - 2011-2018.