| 刘含若,白玮玲,余 双,张 悦,杜一帆,王宁利.眼底图像深度学习技术对屈光不正的智能诊断研究[J].,2020,(18):3587-3591 |
| 眼底图像深度学习技术对屈光不正的智能诊断研究 |
| Research on Intelligent Diagnosis of Refractive Error by Deep Learning Technology of Fundus Images |
| 投稿时间：2020-03-20 修订日期：2020-04-15 |
| DOI：10.13241/j.cnki.pmb.2020.18.042 |
| 中文关键词: 深度学习 屈光不正 智能诊断 |
| 英文关键词: Deep learning Refractive error Intelligent diagnosis |
| 基金项目:国家自然科学基金项目（81700813）；北京市医院管理局"青苗"计划专项经费（QML20180205）；北京市优秀人才培养资助项目；北京市科技新星项目（Z191100001119072）；首都医科大学附属北京同仁医院拔尖人才培养计划，医药协同科研创新研究专项（Z181100001918035）；广东省重点领域研发计划项目（2018B010111001）；深圳市科技计划项目（ZDSYS201802021814180） |
|
| 摘要点击次数: 875 |
| 全文下载次数: 456 |
| 中文摘要: |
| 摘要 目的：提出一种基于人类视觉注意力机制的RE-Net网络结构以使卷积神经网络(CNN)更适用于眼底相的屈光不正的智能诊断评估。方法：RE-Net由ResNet34作为骨干网络，进一步使用了上下文注意力模块，包括通道注意力机制和空间注意力机制，使其相应的通道发挥最大的作用，提高响应区域的权重。结果：使用了4358张眼底图像作为RE-Net的训练集。在包含485张眼底图像的测试集上,分类准确率分别为，高度近视93.3%，中度近视89.7%，轻度近视83.2%，轻度远视82.5%，中度远视79.5%，重度远视84.6%，平均分类准确率达85.5%, 曲线下面积(AUC)为0.909，灵敏度为0. 93, 特异性为0. 89, Kappa值为0. 79 (χ2 ＝23.21，P＜0. 05)。结论：基于深度学习的RE-NET人工智能诊断系统能较好进行屈光不正的诊断评估,有望为屈光不正提供一种新的筛查工具。 |
| 英文摘要: |
| ABSTRACT Objective: To propose a RE-Net network structure based on the human visual attention mechanism to make the convolutional neural network (CNN) more suitable for the intelligent diagnosis and evaluation of fundus refractive errors. Methods: In this study, ResNet34 was used as the backbone network, and further used the contextual attention module, including channel attention mechanism and spatial attention mechanism, so that the corresponding channel could play the maximum role and increase the weight of the response area. Results: 4358 fundus images were used as the training set of RE-Net. On the test set containing 485 fundus images, the classification accuracy rates were 93.3% for high myopia, 89.7% for moderate myopia, 83.2% for mild myopia, 82.5% for mild hyperopia, 79.5% for moderate hyperopia, and 84.6% for severe hyperopia. The average classification accuracy rate was 85.5%, and the area under the curve (AUC) was 0.909, with sensitivity 0.93, specificity 0.89, and the Kappa value was 0.79 (χ2=23.21, P<0.05). Conclusion: The RE-NET artificial intelligence diagnosis system based on deep learning can better diagnose and evaluate refractive errors and is expected to provide a new screening tool for refractive errors. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
