Computer Vision News - July 2018

Juan explains how they do this: “ We are using convolutional neural networks, which is the mainstream technique to analyse and process images nowadays. We take the images and those images usually have multiple cells, so we use neural networks to extract each individual cell from the image. Then we train another network which learns to recognise the differences with respect to other cells. That’s the thing that we use in order to differentiate if a treatment works or if it doesn’t. ” The work that Juan is presenting is the first time that they have used neural networks and convolutional networks for the analysis of this type of biological image. They are using it to understand cancer mutations. In the poster, he will present some examples of how they detect when a mutation has a difference with respect to healthy cells. Juan explains that one of the main problems in cancer treatment nowadays is that a single tumour can have multiple mutations. We don’t know exactly which of those mutations is the one that drives the cancer. We can identify all of those mutations, but we don’t know which one is the problematic one. With this image analysis technique, they can detect which of the mutations are having an impact in the growing of the cancer or the type of cells that are present. He believes that with this they can unlock the difficulties of treating cancer at a much larger scale. Thinking about next steps, Juan says that right now they are analysing even more mutations. It’s something that was not possible before, because the techniques to understand when a mutation is impactful or not cannot be run at larger scale. It’s the power of images and he says that’s why they are presenting their work at CVPR, because they can run a larger scale study with millions of images in order to understand pretty much any mutation in the human genome. He hopes they will scale it up to many more mutations in order to prevent and treat cancer. Juan presented his poster on June 21. Thursday 41 Juan Caicedo