Mitigating annotation shift in cancer classification using single image generative models