深度学习在糖尿病视网膜病灶检测中的应用综述

Review of Application of Deep Learning in Detection of Diabetic Retinal Lesions

糖尿病视网膜病变是世界上致盲率最高的眼科疾病,早期诊断可以显著降低患者失明的概率。深度学习方法可以提取医学图像的隐含特征,并完成图像的检测任务,因此应用深度学习实现糖尿病视网膜病灶检测成为研究热点。主要从数据集介绍、全监督检测方法、非完全监督检测方法、小样本问题的处理和模型可解释性五个方面进行详细总结,重点整理各类方法的基本思想、网络结构形式、改进方案及优缺点总结等内容,结合当前检测方法所面临的挑战,对其未来研究方向进行展望。

Diabetic retinopathy is the eye disease with the highest blindness rate in the world,and early diagnosis can significantly reduce the probability of blindness in patients.Deep learning methods can extract hidden features of medical images and complete image detection tasks.Therefore,the application of deep learning to detect diabetic retinal lesions has become a research hotspot.This article mainly summarizes five aspects:data set introduction,fully-supervised detection methods,non-fully-supervised detection methods,handling of small sample problems and model interpretability.It focuses on sorting out the basic ideas of various methods,network structure forms,improvement plans and summaries of advantages and disadvantages,etc.Finally,combined with the challenges faced by current detection methods,the future research directions are analyzed.