MIDL Vision 2022

In beginning to address this, Alexander and the team tried existing state-of-the-art domain adaptation methods, but none were suitable for this problem. They set to work finding a better solution. “ Our goal was to guide the learning process on the unlabeled data of covered patients with some kind of direct supervision , ” he explains. “ We needed supervision that was definitely correct and not noisy like pseudo labels, for example. The idea is based on our observation that humans can intuitively judge whether a pose predicted by a model is a valid pose that a human can take or if it’s anatomically implausible. The main idea is to provide this prior anatomical knowledge to the model as a loss function. ” Alexander tells us the next step for this work will be to directly embed this prior anatomical knowledge into the network architecture itself , such that implausible predictions are prevented by network design. He hopes this will improve the model’s robustness and generalization. “ An important problem of deep learning is the necessity of labels, and also the problem of generalization under domain shifts, ” he adds. “ I have a keen interest in domain adaptation and generalization . It’s the problem we’ve addressed in this work, and I found it fascinating to learn to see under the cover without direct supervision. ” 13 Alexander Bigalke VISION MIDL

RkJQdWJsaXNoZXIy NTc3NzU=