Medical image segmentation: going beyond standard CNNs

Abstract: Medical image segmentation serves as a key step for the detection, diagnosis and tracking of various diseases. While deep learning techniques such as convolutional neural networks (CNNs) have achieved outstanding performances for this task, they typically require large amounts of expert-annotated data which are rarely available in real-life applications. This talk will present recent advances in medical image segmentation that extend traditional models like CNNs in three different ways. First, we will see how unlabeled data and constrained optimization can be leveraged during training to improve the learning of model parameters. We will then describe how adversarial learning can overcome the problem of labeled data scarcity using both unlabeled and generated images in training. Lastly, we will present a recent application of graph-based CNNs to cortical surface parcellation.

 

Speaker Bio: Christian Desrosiers is working since 2009 as assistant professor in the Software and IT Engineering department at ÉTS. Before joining the departement, he was a postdoctoral research assistant at the University of Minnesota under the supervision of professor George Karypis. Christian Desrosiers obtained his Ph.D. in applied mathematics at École Polytechnique de Montréal in 2008. His main areas of research are data mining machine learning biomedical imaging recommender systems and business intelligence.