Computer Vision News - July 2019

Amy Zhao is a graduate student at MIT in the final year of her PhD. She spoke to us ahead of her oral and poster: Data Augmentation Using Learned Transformations for One-Shot Medical Image Segmentation, which explores how to do smarter and better data augmentation to help improve medical image segmentation when you don’t have a lot of labelled data. Amy says that deep learning has been demonstrated to be so powerful for so many image tasks, but it’s very hard to do it on medical data. This work explores segmentation when you only have one labelled example. She says this is a pretty common scenario. It’s easy to get hundreds of unlabelled examples if you need to – MRI scans from patients, for example – but how do we leverage the information in these unlabelled scans to help us with segmenting any scan? These unlabelled scans have a lot of information in them. They show you the anatomical variations in your population. They’ll also have intensity variations because they were taken with different scanners and they’re of different patients. This method can learn to mimic all of these variations and synthesise examples that look like these unlabelled scans. These examples can then be used to train a segmentation CNN to leverage the power of deep learning. Amy tells us that there’s a more common way of doing what they’re trying to do: “ Data augmentation. Most people who are familiar with computer vision know how to use it to some extent. For medical imaging, it’s common to do data augmentation using random smooth flow fields. These are fairly easy to code, but they don’t always produce the best results and you need to hand tune the parameters. The challenge here is that medical imaging is a pretty complex space. These brain MRIs have very complex variations. It’s difficult for people to write functions that mimic these variations. We are trying to learn to do this instead of trying to do it by hand. ” Data Augmentation Using Learned Transformations for One-Shot Medical Image Segmentation 26 DAILY CVPR Thursday Presentation “ These brain MRIs have very complex variations. It’s difficult for people to write functions that mimic these variations. We are trying to learn to do this instead of trying to do it by hand. ”