Computer Vision News - August 2020

using some labels to make the encoder more powerful. The images are massive, but by using the encoder and compressing these big images into something that has these important features, they are already solving the vision task for the machine. The machine can then focus on other things like predicting outcomes for patients. It’s separating the problem into two. First, it solves the vision, so the raw pixels are no longer needed, and second, it targets labels for the patient . Being able to train on entire images will lead to better outcomes for patients. “Suddenly we have something which represents the patients that we can train our model on,” he says enthusiastically. “Using biomarker data and markers that are available in oncology , we can predict, for instance, whether a certain therapy is going to be good for a patient or not, or whether a tumor will be more aggressive in the future or not. We haven’t had any other method to tackle this before.” The work is still at the research stage and it will take some time to finetune it before it is ready to proceed towards a real-world clinical application . He thinks it is important to emphasize this, as people who are sick often search for new technologies and therapies that could help. Finding novel research like this can sometimes provide false hope. One thing he thinks would move the process along is if there was more David Tellez 21 Best of MIDL 2020 "Neural networks work differently. They see numbers and have to make sense of the pixels that are in the images."