| Summary: | One prominent issue in the application of deep learning is the failure to generalize to data that lies on a different distribution to the training data. While many methods have been proposed to address this, prior work has shown that when operating under the same conditions most algorithms perform almost equally. As such, more work needs to be done to validate past and future methods before they are put into important scenarios like medical imaging. Our work analyses eight domain generalization algorithms across four important medical imaging classification datasets along with three standard natural image classification problems to discover the differences in how these methods operate in these different contexts. We assess these algorithms in terms of generalization capability, domain invariance, and representational sensitivity. Through this, we show that despite the differences between domain and content variations between natural and medical imaging there is little deviation in the operation of each method between natural images and medical images. Additionally, we show that all tested algorithms retain significant amounts of domain-specific information in their feature representations despite explicit training to remove it. Thus, revealing the failure point of all these methods is a lack of class-discriminative features extracted from out-of-distribution data. While these results show that methods that work well on natural imaging work similarly in medical imaging, no method outperforms baseline methods, highlighting the continuing gap of achieving adequate domain generalization. Similarly, the results also question the efficacy of optimizing for domain invariant representations as a method for generalizing to unseen domains.
|
|---|