Deep learning techniques have demonstrated remarkable success in the field of medical image analysis. However, the existence of label noise within data significantly hampers its performance. In this paper, we introduce a novel noise-robust learning method which integrates noise rate estimation into sample selection approaches for handling noisy datasets. We first estimate the noise rate of a dataset with Linear Regression based on the distribution of loss values. Then, potentially noisy samples are excluded based on this estimated noise rate, and sparse regularization is further employed to improve the robustness of our deep learning model. Our proposed method is evaluated on five benchmark medical image classification datasets, including two datasets featuring 3D medical images. Experiments show that our method outperforms other existing noise-robust learning methods, especially when noise rate is very big.