COVID-19 폐 병변 분할 네트워크를 위한 소스-프리 비지도 도메인 적응

The coronavirus disease (COVID-19) has been spread worldwide, causing over 4.55M deaths. The chest CT scan can act as a complimentary tool to the Reverse Transcription Polymerase Chain Reaction (RT-PCR) for diagnosis. The Deep learning (DL)-based networks can automate the diagnosis process by predicting the probability of being positive to the disease and providing lesion segmentation maps. However, these networks are inherently biased to the distributions of training data (source domain) and vulnerable to data from unseen distributions (target domain). To overcome the impracticality, we propose a novel source-free unsupervised adaptation (UDA) method for COVID-19 lesion segmentation networks, which distills knowledge from target domain without using manual annotations. It uses predicted segmentation maps after weighted label correction (WLC) as pseudo labels. The experimental results show that the proposed UDA method can recover the performance degradation after applying a source network directly to target domain.