MCSC Almost Done Seminar

Speaker: Eric Ng
Title: A Deep Learning Approach to Discontinuity-Preserving Image Registration
Abstract: Image registration, the task of aligning two images of the same object at different times, has been the subject of research for many years due to its vast applications from medical imaging to computer vision. Traditional image registration involves solving an optimization problem where the objective function is composed of a similarity/dissimilarity measure, followed by a regularizer. In many cases, gradient based regularizers such as diffusion and total variation are applied to enforce smoothness in the transformation field. Since these methods assume global smoothness within the image domain, they pose issues for scenarios where local discontinuities are expected, such as the sliding motion between the lungs and the chest wall during the respiratory cycle. Furthermore, the objective function must be optimized for each image pair, thus registering multiple sets of images become very time consuming and scale poorly to higher resolution image volumes.
Using recent advances in deep learning, we propose an unsupervised learning-based model, where the registration process is formulated as a mapping between image pairs and a deformation field. The model is trained over a loss function with a custom regularizer that preserves local discontinuities, while simultaneously respecting the smoothness assumption in homogeneous regions of image volumes. The model is evaluated using publicly available 4D CT datasets from DIR-Lab and the Point-validated Pixel-based Breathing Thorax (POPI) model.

Speaker: Rory Coles
Title: Using Machine Learning Methods to Aid Scientists in Laboratory Environments
Abstract: As machine learning gains popularity as a scientific instrument, we look to create methods to implement it as a laboratory tool for researchers. In the first of two projects, we discuss creating a real-time signal interference monitor for use at a radio observatory. We show how deep neural networks can be used to assist with the detection of radio-frequency interference around the site, and consider methods of unsupervised learning to identify patterns in the detections.
In the second project, we show how a reinforcement learning agent can build an internal hypothesis of its environment, using experience from past measurements, that it can then act on. We demonstrate how our newly developed method can be used to learn the dynamics of physics-based models and exploit the knowledge gained to achieve a given objective with measurable confidence. We also demonstrate how the agent’s behaviour changes when the frequency of certain measurements is limited.