| With the development of deep learning,significant improvements have been made in medical image classification technology.However,deep learning requires a large amount of labeled data.As medical experts find labeling samples time-consuming and laborious,and collecting labels from crowdsourcing may introduce noisy labels.Deep Neural Networks(DNNs)tend to overfit to noisy label data,which is detrimental to training DNNs and reduces the accuracy of classifiers.Therefore,there is an urgent need for effective methods to handle noisy labels.However,there is a lack of research on handling noisy labels in deep learning for medical images.To address this issue,it is crucial to mitigate the impact of these noisy labels,and many existing methods attempt to classify training data into clean and noisy datasets based on loss values,followed by different treatments of noisy-label data.The key to improving model performance is to deal with hard samples,but regardless of whether the label is clean or noisy,hard samples tend to have relatively larger losses.Therefore,current methods cannot accurately classify hard samples.This paper investigates the aforementioned issues and proposes the following research findings.(1)This thesis proposes a new sample classification criterion that preliminarily classifies training samples into clean,hard,and noisy samples based on whether the predictions of two networks are consistent and whether the prediction results of the network match the given label.Different treatments are applied to different types of samples,enabling the model to be effectively trained on noisy datasets.(2)This thesis proposes a sample detection scheme that further accurately classifies clean,hard,and noisy samples based on preliminary classification.The detected samples are then integrated into the label correction stage of this paper,which can make the original dataset cleaner while retaining more hard samples.(3)After obtaining a cleaner dataset,this thesis further proposes a collaborative learning scheme to train a noise-robust classification model.More noisy samples are discarded to reduce the harm of noisy labels,while maximizing the value of hard samples.Extensive experiments show that this algorithm can not only more accurately correct erroneous labels in the original dataset but also achieve robust classification against noisy label interference on artificial and real noisy datasets.Compared with state-of-the-art methods,this algorithm has made substantial improvements. |
