基于增强型半监督生成对抗网络的糖尿病视网膜病变识别

Diabetic Retinopathy Recognition Using an Enhanced Semi-Supervised Generative Adversarial Networks

糖尿病视网膜病变(Diabetic Retinopathy, DR)是由糖尿病引起的视网膜血管壁受损致使视觉功能下降的一种具有特异性改变的眼底病变,是主要致盲疾病之一。在医学图像处理中,糖尿病视网膜病变诊疗通常面临高质量标注样本少和未标注数据不能充分利用的困境。基于此,本文利用增强的半监督生成对抗网络对糖尿病视网膜病变等级和程度进行识别,实现更高的识别精度和泛化能力,最终四分类任务中准确率达到77.2%,二分类任务中AUC达到93.9%。

As one of the main blinding diseases, Diabetic Retinopathy (DR) is a kind of specific fundus lesion with specific changes in visual function caused by damage to the retinal vessel wall caused by dia-betes. In medical image processing, the treatment of DR usually faces the dilemma of lacked of high-quality labeled samples and unlabeled data that cannot be fully utilized. Based on that, in this paper, the enhanced semi-supervised generative adversarial network is used to identify the grade and extent of DR. Finally, it can achieve higher recognition accuracy and generalization ability, that is, the accuracy rate reaches 77.2% in the four classification task, and the AUC reaches 93.9% in the two classification task.