A progressive framework for data efficiency enhancement in contrastive learning for COVID-19 CT image classification

Despite the initial outbreak of COVID-19 occurring years ago, the continued emergence of SARS-CoV-2 variants poses persistent challenges to global public health systems. Computed tomography (CT) imaging remains an essential tool for detecting lung abnormalities associated with COVID-19. While self-supervised learning has demonstrated significant potential in computer-aided diagnosis (CAD), its application is hindered by the redundancy in unlabeled datasets and the intricate nature of pathological patterns. To address these challenges, we propose a progressive framework for self-supervised contrastive learning in COVID-19 CT image classification. Our approach integrates hash-based redundancy reduction to refine the dataset and a hard negative synthesis strategy to enhance feature learning during pre-training. Experimental results demonstrate that our method achieves notable gains in downstream classification performance, highlighting its potential for medical image analysis tasks.