A Light-Weight Vision Transformer For COVID-19 Chest X-Ray Images Classification

As a global pandemic infectious disease, COVID-19 has caused huge loss of life and economic property to the world. Using deep learning techniques to rapidly diagnose and distinguish positive patients by chest X-ray images can help reduce the healthcare burden. Previous studies have mainly used traditional CNN models, pure Vision Transformer models or direct combination of both, with large parameter operations and average accuracy. In this paper, we propose a hybrid light-weight model architecture based on CNN and Vision Transformer inspired by MobileViT block. Considering the imbalance of existing datasets categories, we apply a novel loss function Large Marginal Perceptual Focal Length Loss (LMFLoss) to enhance the comprehensive performance of the model and its generalization ability. Proposed model achieves State Of The Art (SOTA) performance in terms of total classification accuracy, precision, recall, and F1 score, and for the COVID-19 category, the model obtains 100% on all classification metrics.