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.

本文介绍一种新的噪声鲁棒学习方法，将噪声率估计整合到样本选择方法中，以处理有噪声的数据集，首先根据损失值的分布使用线性回归估计数据集的噪声率，然后根据估计的噪声率排除可能有噪声的样本，并进一步使用稀疏正则化来提高我们的深度学习模型的鲁棒性，实验证明我们的方法在处理具有较大噪声率的数据集时优于其他现有的噪声鲁棒学习方法。

医学图像分析中的噪声学习中，带有噪声率估计的样本选择