基于噪声鲁棒学习的胸部X射线尘肺鉴别方法

Method of Noise Robust Learning Based Chest X-ray Discrimination of Pneumoconiosis

在医学图像噪声标注数据的训练中,目前常用的方法是根据训练损失对噪声标签数据集进行划分,以过滤掉噪声标签样本。然而,这种方法面临两个需要解决的问题,即如何在筛选出噪声样本的同时尽可能地保留与其损失分布相似的困难样本,以及如何提高样本利用率,挖掘隐藏在噪声样本中的有用信息以减轻模型过拟合的问题。为了解决上述问题,提出一种由样本分布引导的噪声鲁棒学习策略(SGRL),包括样本划分与半监督对比分类。为了更可靠地区分信息量大的困难样本与有害噪声样本,介绍一种噪声滤波器样本选择方法。此外,提出了一种增强匹配对比网络,使用所有样本进行训练,从而得到一个具有噪声鲁棒性的分类模型。在此基础上,利用对比学习作为补充,进一步对抗对噪声标签的记忆,提高筛查准确率。实验结果表明,该方法在5%、10%、20%和40%噪声比的尘肺胸片数据集上均取得了显著的性能提升。与现有的先进方法相比,该方法的筛查准确率分别平均提升了5.88、7.05、7.59和6.19个百分点,验证了改进方法的有效性。

In training medical image noise annotation data,the prevailing approach involves partitioning the noise-labeled dataset based on training loss to filter out the noise-labeled samples.However,this method faces two pressing issues that require resolution:first,filtering out noise samples while retaining difficult samples with similar loss distributions as much as possible,and second,enhancing sample utilization and uncovering valuable information embedded in noise samples to alleviate model overfitting.This study proposes a Sample Distribution Guided Noise Robust Learning strategy(SGRL)comprising sample partitioning and semi-supervised contrastive classification to address these challenges.A straightforward yet effective sample selection method called a noise filter method is introduced to distinguish informative,difficult samples from detrimental noise samples more accurately.Additionally,an enhanced matching contrastive network is proposed to train using all samples,yielding a noise-robust classification model.Contrastive learning is utilized as a supplement to counter the memorization of noise labels and improve screening accuracy.The experimental results demonstrate significant performance improvement of the proposed method across dust-induced pneumoconiosis chest X-ray datasets with noise ratios of 5%,10%,20%,and 40%.Compared with existing state-of-the-art methods,the screening accuracy of this method increased by an average of 5.88,7.05,7.59,and 6.19 percentage points,validating the effectiveness of the proposed improvement method.