Modern deep learning (DL) algorithms rely extensively on large amounts of annotated data. Even when a large dataset is available, DL algorithms often fail miserably when deployed to settings with data characteristics significantly differing from those used for training. Domain adaptation (DA) and domain generalization (DG) algorithms aim to mitigate the gap between source (train) and target (test) distributions by learning domain-agnostic features or minimizing the discrepancy in the model’s predictions between the source and target distributions. This issue is prevalent in practical medical imaging settings, as the cost of obtaining both images and annotations is extremely expensive, limiting data accessibility to only a bulk of images collected from a few hospitals or detector devices, but a model must be suitable for multi-center, multi-device settings. In this seminar, we will cover existing literature on DA and DG, discussing their capabilities, assumptions, methodologies, along with their limitations. The session will conclude with research directions relevant to pragmatic industrial settings.
|
