TL-med: A Two-stage transfer learning recognition model for medical images of COVID-19

•We propose a two-stage transfer learning method, which can effectively learn the information between different modal images.•We use Vision Transformer as the backbone network, which effectively improves the classification accuracy and efficiency.•This method effectively alleviates the problem of insufficient data set.•This method achieves an accuracy of 93.24% in the medical images classification of COVID-19.