| 摘要: |
| 目的 利用深度学习方法自动提取眼底白内障特征，构建白内障自动分类器，并可视化分析深度网络中间层特征的逐层变换过程。方法 基于临床眼底图像，使用深度卷积神经网络（CNN）从输入数据的原始表示直接学习有用的特征，对比分析CNN自动提取的特征与预定义特征的性能表现。然后利用反卷积神经网络（DN）量化分析CNN各个中间层的特征，进一步研究输入图像中对CNN的预测贡献最大的像素集，探究CNN表征白内障的具体过程。结果 使用深度学习方法构建的分类器在四分类任务中达到0.818 6的平均准确率。与现有的预定义特征集相比，利用深度CNN自动提取的特征集能提供更好的白内障特征表示。CNN中间层特征呈现从低级抽象到高级抽象的分层变换，如梯度变化到边缘，然后到边缘状发散结构的组合，最后到血管和视神经盘信息的高级抽象，这种变换过程与临床检测白内障的诊断标准相吻合。结论 基于深度学习的分类器在性能表现上优于现有分类器。该方法对检测其他眼病也可能具有潜在的应用前景。 |
| 关键词: 人工智能 白内障 深度学习 深度卷积神经网络 反卷积神经网络 |
| DOI：10.16781/j.0258-879x.2018.08.0878 |
| 投稿时间：2018-06-21修订日期：2018-07-26 |
| 基金项目:国家科技重大专项（2017YFB1400803），国家自然科学基金重点项目（71432004）. |
|
| Cataract recognition and grading based on deep learning |
| LI Jian-qiang1*,ZHANG Ling-lin1,ZHANG Li2,YANG Ji-jiang3,WANG Qing3 |
(1. School of Software Engineering, Beijing University of Technology, Beijing 100124, China;
2. Beijing Tongren Eye Center, Beijing Tongren Hospital, Capital Medical University, Beijing 100730, China;
3. Research Institute of Information Technology, Tsinghua University, Beijing 100084, China
*Corresponding author) |
| Abstract: |
| Objective To automatically extract the characteristics of fundus cataract by deep learning, construct a automatic classifier for cataract, and visualize the layer-by-layer feature transformation process of the intermediate layer of deep network.Methods Based on the clinical fundus image, a deep convolutional neural network (CNN) was used to directly learn useful features from the original representation of input data, and then the features extracted by the CNN were compared with pre-defined features. The deconvolution neural network (DN) method was used to quantitatively analyze the characteristics of each intermediate layer of CNN, analyze the pixel sets that have the most contribution to the prediction performance of CNN in the input image, and explore the process in characterizing cataract by CNN.Results The classifier constructed by deep learning achieved an average accuracy of 0.818 6 in four-category tasks. Compared with the existing pre-defined feature set, the feature set automatically extracted by the deep CNN performed better in representing characteristics of cataract. The features of the intermediate layer of CNN hierarchically transformed from low-level abstraction to high-level abstraction, including changed from gradient to edge, then to the combination of edge-like divergent structures, and finally to the high-level abstraction of blood vessel and optic disc information, and this transformation process coincided with the clinical diagnostic criteria of cataract.Conclusion The classifier based on deep learning is superior to the existing classifier in terms of performance. In addition, this method has potential application in detecting other eye diseases. |
| Key words: artificial intelligence cataract deep learning deep convolutional neural network deconvolution neural network |
