摘要
深度卷积神经网络(deep convolutional neural networks,DCNN)模型因其高效的学习表征能力,被广泛应用于包括肺病等各类型疾病的辅助诊断。现有DCNN模型大都具有自监督性,对数据集的数量和质量要求较高。然而,医疗数据的隐私性和数据标注的成本极大限制了模型的性能和效力。对此,提出一种基于对比学习的阶段式自监督诊断模型。该模型对随机数据增强后的未标记肺部CT图像进行特征提取,且不需要过多对标注COVID-19数据集进行训练,有效解决了因标注数据集少而产生有监督模型训练困难的问题。同时,设计了4个消融实验来验证模型的性能。实验结果表明,提出的模型在新冠肺炎CT图像自监督诊断任务中有良好的表现。
Because of its efficient learning and characterization capabilities,the deep convolutional neural network(DCNN)model is widely used in the auxiliary diagnosis of various types of diseases,including lung disease.Most of the existing DCNN models are self-supervised,which have high requirements for the quantity and quality of data sets.However,the privacy of medical data and the high cost of data labeling greatly limit the performance and effectiveness of the model.A phased self-superivised diagnostic model based on comparative learning was proposed.The model extracted the features of unlabeled lung CT images augmented by random data which did not require too many labeled COVID-19 data sets for training,which effectively solved the problem that the small labeled data sets made it difficult to train supervised models.At the same time,four ablation experiments were designed to verify the performance of this model.The experimental results show that the model has excellent performance in the self-supervised diagnosis task of lung CT images of COVID-19.
作者
王栩飞
齐义文
岑游彬
郭诗彤
李贺
WANG Xufei;QI Yiwen;CEN Youbin;GUO Shitong;LI He(College of Automation,Shenyang Aerospace University,Shenyang 110136,China)
出处
《沈阳航空航天大学学报》
2023年第6期59-67,共9页
Journal of Shenyang Aerospace University
基金
国家自然科学基金(项目编号:62003223)
中央引导地方科技发展专项(项目编号:2021JH6/10500162)。
