MIDL Vision 2021

to find a good alignment? Much of this work at the time was unsupervised, but we already had nicely registered examples, so we used supervised learning to generate image pairs by mapping the initial image to the 2p image domain so that it appears like a 2p image. We wanted to include more data that hadn’t been registered, so we used an additional unsupervised discriminator loss. ” In GAN , the discriminator distinguishes between a real image and a generated image . The network that maps the image is trying to create better images, so they play them off against each other. In this case, the generator tries to be better so that the discriminator cannot distinguish between the image pairs that have been registered and the images without ground truth. “ This kind of image-to-image translation is often used in image segmentation tasks , ” Henrik points out. “ For example, where you have a model that can detect a tumor in a T1 MRI image, but you want to apply your segmentation method to T2 MRI 16 Presentation VISION MIDL The Nissl and Backlit images are part of our Brain/MINDS marmoset brain image database. Above, images of neural tracers, Nissl, backlit as well as atlas information in the same image space.