Spectrum and Style Transformation Framework for Omni-Domain COVID-19 Diagnosis

Abstract

Coronavirus disease 2019 (COVID-19) is an ongoing global pandemic and profoundly affects almost all people around the world. Thus, many automatic diagnosis methods based on computed tomography (CT) images have been proposed to reduce the workload of radiologists. Most of the existing methods focus on the in-domain predictions, i.e., the training and testing have similar distributions, which is impractical in real-world situations, since the CT images can be collected from different devices and in different hospitals. To improve the diagnosis performance of COVID-19 for both in-domain and out-of-domain data, this paper proposes a spectrum and style transformation framework for omni-domain COVID-19 diagnosis. To achieve this, we first present a spectrum transform module, which helps to discover the discriminating features of each domain to recognize the in-domain data. Then, we formulate a cross-domain stylization module, which learns the cross-domain knowledge to enhance the model generalization capability to deal with out-of-domain data. Moreover, our framework is a plug-and-play module that can be easily integrated into existing deep models. We evaluate our framework on four COVID-19 datasets and show our method consistently improves the diagnosis performance of various methods on both in-domain and out-of-domain data.