The Co-teaching-based sample selection method is popular in learning with noisy labels, which trains two deep networks simultaneously for sample selection based on small-loss criterion. However, the loss is generated based on the networks trained on noisy data, the network cannot guarantee whether the predictions of networks can be trusted or not. In this paper, we propose a robust learning framework called TL (trustworthy learning), TL first estimates uncertainty from label information and evidence provided by two diverged networks. Then, TL employs the dual-criteria (small-loss and uncertainty estimation) to select high confidence samples. Finally, TL designs a weighted joint loss to mitigate the effect of residual noise in the networks. Comprehensive experimental results on both synthetic and real-world noise verify the effectiveness of the proposed method in terms of reliability and robustness for learning with label noise. The code could be accessed from https://github.com/lpyxy/tl .