MIDL Vision 2022

be labeled as a head, but it’s a knee MRI. We’ve found that in our hospital dataset, so we can’t rely on labels. We must assure ourselves that our networks will either only work on the data they’re presented with or, when presented with incorrect data, will work robustly. We don’t want them to look at a knee MRI and say you’ve got a head tumor and need urgent clinical care, because that’s not a desirable or useful outcome. ” There are many challenges with generative modeling and anomaly detection outside of MRI . Mark says papers often aim to build generative models and use the likelihood of those generative models to flag data as in or out of distribution, but these models tend to fail. If you train an algorithm on a dataset of cityscape scenes, for example, and then give it a completely blank image, it will say this is a very high likelihood and completely in distribution. It is always possible to find these failure cases where the model is convinced an image is in distribution when it is evident to a human that it is entirely out of distribution. It is not easy to build robust models . “ Transformer architectures are becoming very big, ” Mark comments. 5 Mark Graham VISION MIDL