ECCV 2020 Daily - Wednesday

of modality representations, better methods of unsupervised and semi-supervised learning, more careful design of datasets, more consideration of edge cases, and more sophisticated model reasoning capabilities. 3 CV4MI 2 DAILY W e d n e s d a y 6 Spatial image conditioning on metadata (input z) of feature maps: the auxiliary network learns to amplify or suppress regions of interest in the image according to information in the vector z [G. Jacenkow et al, due to be presented at MICCAI 2020]. Future opportunities in medical image analysis In the future, we will need to get better at doing the tasks that we know machine learning is capable of (in particular object detection) and build the trust with medical users that our systems are safe and transparent. Ideally algorithms will know when they don’t get it right, and when they do get it right, will be able to explain their results. Next we want to tackle the more difficult tasks that we know AI cannot currently solve very well: anomaly detection, zero-/one-/few-shot learning, multimodal tasks, and more complex tasks requiring reasoning. Ultimately human-machine combinations should be very powerful. Computers don’t get tired, they have a high work rate, and happily they make different types of mistakes to humans. It will be important to keep the domain expert at the heart of everything that we do, and to ensure we develop useful and usable tools that support rather than complicate the clinical workflow. Looking to the longer term, precision medicine is the holy grail i.e. being able to model disease (and health) for each individual, tailored to their genetics and lifestyle. This is going to require learning with much weaker long-range or delayed reward type supervision, and deeper understanding of causality will likely be important. There is some exciting work emerging – so watch this space! Or in the more immediate term, watch Alison’s short talk.