基于欠采样结合半监督学习的肺损伤评级分类OA
Research on Lung Injury Rating Classification Based on Under-Sampling and Semi-Supervised Learning
肺超声可以通过直接或间接征象辅助医生评估肺部病变情况,快速筛查急性呼吸困难病因,有利于医生更好地评估和管理肺部病变患者.文中提出了一种新的自动肺超声评分方法以实现更精准的肺损伤评估.采用随机欠采样预处理解决数据集中存在的类别不平衡问题.通过采用半监督学习方法来更好地利用剔除的样本信息,在训练过程中使用交叉伪损失估计(Cross Pseudo-loss Estimation,CPLE)来选择可靠的伪标记数据进行训练以进一步提升模型性能.实验结果表明,所提方法在实测数据集具有更好的性能,准确率为78.51％,提高了少数类别的分类准确率.
Lung ultrasound can assist doctors in evaluating pulmonary lesions through direct or indirect signs,and quickly screen the causes of acute dyspnea,which is beneficial for doctors to better evaluate and manage patients with pulmonary lesions.This study proposes a new automatic lung ultrasound scoring method to achieve a more accurate assessment of lung injury.Random under-sampling preprocessing is adopted to address the class imbalance problem ex-isting in the dataset.A semi-supervised learning method is employed to make better use of the information of the exclu-ded samples.During the training process,CPLE(Cross Pseudo-loss Estimation)is used to select reliable pseudo-la-beled data for training,further improving the performance of the model.The experimental results show that the pro-posed method has better performance on the measured dataset,with an accuracy rate of 78.51％,and improves the classification accuracy of the minority classes.
黄秋城;张鞠成;黄天海;褚永华;蒋明峰
浙江理工大学计算机科学与技术学院,浙江 杭州 310018浙江大学医学院附属第二医院临床工程部,浙江 杭州 310003浙江大学医学院附属第二医院临床工程部,浙江 杭州 310003浙江大学医学院附属第二医院临床工程部,浙江 杭州 310003浙江理工大学计算机科学与技术学院,浙江 杭州 310018
信息技术与安全科学
肺超声肺损伤评估深度学习图像分类欠采样半监督学习交叉伪监督类别不平衡
lung ultrasoundlung injury assessmentdeep learningimage classificationunder-samplingsemi-supervised learningcross-pseudo-supervisionclass imbalance
《电子科技》 2026 (1)
40-46,7
浙江省科技厅重点研发项目(2023C03088)Key Research and Development Project of Zhejiang Provincial Science and Technology Department(2023C03088)
评论