Computer Vision News - August 2021

MIDL Runner Up 20 Semantic Similarity Metrics for Learned Image Registration Steffen Czolbe is a PhD student in the Department of Computer Science at the University of Copenhagen, under the supervision of Aasa Feragen and Oswin Krause. He presented his work proposing a data-driven semantic similaritymetric for image registration at MIDL 2021 last month, and the judges were so impressed with it that he took home the Best Paper - Runner- up award! He tells us more about his exciting research. Most recent progress in image registration has been achieved via deep learning methods. Modern research is much more likely to train a neural network to predict a transformation that can be applied to register two images, than it is to iteratively optimize their similarity . Comparing two images is a complex task . One simple way to do it is to compare your aligned image to your target image by looking at individual pixel intensities using mean squared error. However, in this paper, Steffen proposes a better way. “When we align pixel intensities we ultimately say as long as the image intensities are the same between the images, they’re well aligned,” Steffen explains. “Instead, what we should align in image registration is areas of the image that have the same meaning. If you are aligning images of the brain, lungs, or another organ, you want to align the same organs and same parts of organs. That’s how I got the idea of using a more semantic representation of the images and focusing on aligning that . I used representation learning to extract representations of the images and ultimately we used these to judge their similarity.” Steffen had worked previously with methods using variational autoencoders and simple loss Steffen Czolbe "That’s how I got the idea of using a more semantic representation of the images and focusing on aligning that!" Best of M I D L