LIFAT Defense - BIRGUI SEKOU Taibou

Abstract:
Retinal image segmentation (RIS) ( i.e. blood vessel and/or optic disc segmentation) can serve as a cue to diagnose, treat and monitor diseases such as the diabetic retinopathy, the hypertension, the arteriosclerosis, and the glaucoma.
In this thesis, we explored various learning based models on the task of RIS, namely discriminative dictionary learning (DDL) approaches and artificial neural networks (ANNs) ones. In the first part, four DDL-based models are experimented on the task of retinal blood vessel segmentation. The later are obtained from an initial state-of-the-art review that groups the DDL-based models into four categories.
The final results show the effectiveness of the four methods compared with other state-of-the-art (non-ANN based) models. We then turn to study ANN-based models on the task of RIS. The second part is dedicated to deep learning based models which, generally, require a large number of samples for appropriate training, a requirement that is difficult to satisfy in the medical field. This issue can usually be avoided with a proper initialization of the weights: data augmentation and/or transfer learning. We proposed a framework that brings together these two techniques and can train arbitrarily designed networks that segment an image in one forward pass, with a focus on relatively small databases. An experimental work has been carried out on four publicly available databases using three types of network architecture. The final results show the efficiency of the proposed framework along with state-of-the-art results on all the databases.

Keywords: Deep learning, medical images, image segmentation, transfer learning, dictionary learning, discriminative dictionaries