Improving Medical Image Classification in Noisy Labels Using Only Self-supervised Pretraining