基于深度学习的糖尿病视网膜病变计算机辅助诊断技术研究

Diabetic retinopathy (DR) is the earliest organ complication of diabetes mellitus. Early detection and continuous tracking of DR are essential to control the disease in the early period. Formerly, the screening of DR depended on the use of image segmentation technique, which can automatically segment the lesions rather than classify lesions and determine their nature. Along with the application of Deep Learning approaches, some high quality works grade DR using CNN, but they can neither point out the position of lesions, nor determine lesions’ nature. The main contributions of this project are as follows: (1) this project will integrate the two techniques above and simulate the way ophthalmologists interpret fundus images, which uses image segmentation technique to segment fundus images and uses Deep learning approach to determine the nature of lesions for automatic diagnosis of DR. (2) This project, corresponding to transfer learning, will train the model with 88000 fundus images from EyePacs and validate the model with labelled DR dataset from Tongren Hospital. (3) A Convolutional Neural Network will be established and it will be trained by fundus images labelled with exudations and hemorrhages by Tongren Hospital. This network will screen lesions automatically and get a better result of automatic image registration, so that it will be easy to measure the lesions quantificationally. This project will provide a methodology support for monitoring the disease and provide a scientific basis for DR early diagnosis and treatment.

糖尿病视网膜病变（DR）是糖尿病最早发生的器官并发症，尽早发现与连续跟踪DR是早期防治的关键。DR识别以往主要依靠图像分割技术，这类技术通常智能分割病灶，不能深入判断病灶性质，细分归类病症。随着深度学习在DR领域的应用，出现了几篇高质量的DR分级论文。但目前深度学习只能给出DR分级，不能指出病变所在，也不能鉴别病灶性质。本项目主要贡献如下：1、整合上述两种技术，模拟眼科专家阅片模式，先利用图像分割技术筛选病灶，再利用深度学习技术进行病灶性质细分及智能诊断；2、采用迁移学习方法，在EyePacs上已有88000张眼底影像上训练模型，然后在同仁医院数据库已有标注DR数据集上测试；3、使用深度学习中卷积神经网络，在同仁医院已有标注上实现出血、渗出等病灶的自动识别，并结合不同时期图像的自动配准结果定量描述该区域的变化情况。本项目研究成果能够为DR 早期诊断、治疗提供科学依据，监控病情发展。

糖尿病视网膜病变（DR）是糖尿病最早发生的器官并发症，尽早发现与连续跟踪DR是早期防治的关键。DR识别以往主要依靠图像分割技术，这类技术通常智能分割病灶，不能深入判断病灶性质，细分归类病症。本项目模拟眼科专家阅片模式，先利用图像分割技术筛选病灶，再利用深度学习技术进行病灶性质细分及智能诊断；使用深度学习中卷积神经网络，在已有标注上实现出血、渗出等病灶的自动识别，并结合不同时期图像的自动配准结果定量描述该区域的变化情况。结果显示糖尿病视网膜病变5分类准确率0.874，kappa值0.922，二分类敏感性0.95，特异性0.98（Google为0.90，0.99）。本项目研究成果能够为DR 早期诊断、治疗提供科学依据，监控病情发展。