ADS Deep Dive: Deep Learning in Medical Imaging
Introduction & Chair: Max Welling, Research Chair in Machine Learning, Informatics Institute UvA
We will have 3 speakers, each presenting for 30 minutes plus 10 minutes discussion time:
14:00-14:40: Jörn-Henrik Jacobsen, Informatics Institute UvA on “Inverting Deep Networks – The Why, How and Implications”
Inverting deep neural networks has been proposed as a potential tool to investigate the information discarded with depth. One way to obtain such inverses is to make use of priors that regularize the inversion. However, such priors are typically based on simple heuristics, or come in the form of inflexible deep networks themselves.
I will discuss how to invert arbitrary deep networks with simple learned priors that are useful for other downstream tasks as well. Further, I will introduce a class of deep networks that are invertible by construction and highlight the implications of such architectures.
14:40-15:20: Matthan Caan, Brain Imaging Center, Academic Medical Center (AMC) on “Deep Learning for accelerated MRI reconstruction”
High resolution Magnetic Resonance Imaging (MRI) of the human brain is a timely procedure. In order to accelerate MRI, k-space can be incoherently undersampled beyond the Nyquist-criterion. Compressed Sensing (CS) reconstructs images using a predefined transform, such as the Wavelet transform. Deep learning applies multi-layered neural networks as universal function approximators and is able to find its own compression implicitly. This allows to further accelerate image acquisition. We propose a Recurrent Inference Machine (RIM), which can acquire great network depth, while retaining a low number of parameters. We demonstrate its performance on accelerating MRI-scans of the human brain.
15:20-16:00: Joost Batenburg, Scientific Staff Member & Group Leader, Department of Computational Imaging CWI and Professor, Leiden University on “Real-Time 3D Tomography”
16:00: Coffee, snack & networking