MIDL Vision 2021

scans . That’s why we’re the first to do this because other studies have tried to make a triage model using around 1,000 or less carefully labelled scans, but that’s just not enough for deep learning. ” Other methods have tried to use research databases with a few thousands labeled scans, but David says that when you use these in a real-world clinical setting they fail. The research scans are often unrealistically good , whereas in a hospital, everything is done very quickly, with low dimensions and movement artefacts. To get a model that is going to work, you need to train it on images from that hospital. David’s paper has managed to snatch one of the hotly contested long oral discussion slots at MIDL this year. What does he think separates it from the rest? “ There’s a tendency in medical machine learning to just use things like accuracy, area under the receiver operating curve, sensitivity, and things like that, which are all fine, but what you really need to do is show the clinical impact that something is going to have , ” he explains. “ You can say something is 99% accurate and that sounds great, but if there’s only a 0.01% chance of getting the disease, then it’s easy to get to 99% and just say everything is normal! Sometimes those things are misleading . ” 8 Presentation VISION MIDL