Computer Vision News - August 2020

David Tellez is a PhD student in the Computational Pathology Group at Radboud University in Nijmegen under the supervision of Professor Jeroen van der Laak. He speaks to us about his recent work, which scooped a runner-up prize in the Best Paper category at MIDL this year. David works in a field called computational pathology, which supports pathologists to improve patient care using artificial intelligence. Pathologists are doctors who diagnose disease by examining cells and tissues. For instance, when a patient with cancer has a biopsy taken of their tumor, a pathologist will look at it under a microscope and interpret how the cells are reacting . As humans, we look at the world and recognize familiar patterns. That concept led David to wonder if something similar could be applied in pathology. When pathologists look at these medical images, they see patterns too. They immediately recognize cells and structures and clusters anddifferent types of tissue. Neural networks work differently . They see numbers and have to make sense of the pixels that are in the images. In his paper, David proposes a model where a neural network is trained to mimic a human pathologist , by looking only for interesting features, and extracting them from the pixels. The problem is the images are huge. “If you printed it out on paper one of these images would be the size of a five-story building. You’d need to navigate it like Google Maps!” he jokes. “Extracting these features is not an easy task. You need to train an encoder network that is able to recognize interesting things in the images. It’s computationally very complicated.” David has been inspired by other work he has seen about unsupervised learning . With so many images available, but annotations expensive, what if an encoder could be trained in an unsupervised way to recognize patterns ? He has extended that idea by Extending Unsupervised Neural Image Compression with Supervised Multitask Learning Runner-up for Best Paper Award 20 Best of MIDL 2020