Domain Adaptation Using Pseudo Labels for COVID-19 Detection

Abstract

Deep learning has offered advanced analytical capabilities to enhance the accuracy and efficiency of detecting COVID-19 through complex pattern recognition in medical imaging data. However, the variability across datasets from different domains poses a significant challenge to the generalization abilities of deep learning models. In this paper, we propose a novel two-stage framework for domain adaptation of COVID-19 detection. Initially, We train a model on annotated data from both domains, integrating contrastive representation learning and a modified version of CORAL loss to minimize domain discrepancies. In the subsequent stage, we employ a pseudo-labeling strategy to effectively utilize non-annotated data from the target domain, further enhancing the model’s adaptability and performance. The effectiveness of our approach is demonstrated through extensive experiments, showing significant improvements in COVID-19 detection performance compared to the baseline model. On the COVID-19 domain adaptation leaderboard in the 4th COV19D Competition, our approach ranked 1st with a Macro F1 Score of 77.55%.