基于生成对抗网络的彩色眼底图像硬性渗出检测方法

糖尿病视网膜病变(DR)是糖尿病严重的并发症,是视力损害最常见的原因之一。硬性渗出物(HE)是DR早期的症状之一,从眼底图像中对硬性渗出的准确检测是DR筛查的关键步骤。提出一种基于生成对抗网络(GANs)的视网膜硬性渗出的自动检测方法。相比一般的卷积神经网络,生成对抗网络由生成式模型G和判别式模型D组成,两者之间的博弈与竞争使得生成对抗网络能够更加精确地检测眼底图像中的硬性渗出。首先,为了避免视盘对后续硬性渗出检测的干扰,根据血管分布信息与全局灰度信息,准确定位视盘(OD)中心并掩盖视盘;然后,交替迭代训练生成式模型G和判别式模型D,得到在验证集上分割效果最佳的模型并保存。所提出的算法在e-ophtha EX数据库上训练和验证,并进行像素级评估,获得88.6％、84.3％和86.4％的平均灵敏度、PPV和F-score。在另一个独立的DIARETDB1数据库上进行测试,获得的平均灵敏度、特异性和准确性分别为100％、96.2％和97.8％。综上所述,两个视网膜图像数据库的评估结果证明,生成对抗网络的博弈模式能够有效地检测彩色眼底图像中的硬性渗出。

Hard Exudates Detection Method Based on Generative Adversarial Networks in Color Fundus Images

Diabetic retinopathy (DR) is a serious complication of diabetes and one of the most common causes of visual impairment. Hard exudates (HE) are early symptoms of DR, and its detection plays a key role in screening DR using fundus images. In this study, an automatic computational method based on Generative Adversarial Networks (GANs) for HE detection was proposed and evaluated on a publicly available database (e-ophtha EX). Compared with the general Convolutional Neural Network (CNN),GANs was composed of the generator G and the discriminator D, the mutual game and competition between them made the GANs detect HE in color fundus images more accurately. Firstly, the center of optic disc (OD) was located by considering its vessels features and brightness information, and masked to avoid interference with HE detection. Secondly, the generator G and the discriminator D were trained on e-ophtha EX dataset alternately, with the optimal model determined, which had the best segmentation effect on the validation dataset. The proposed algorithm was validated on e-ophtha EX database on pixel level, achieving the overall average sensitivity, PPV and F-score of 88.6%, 84.3% and 86.4%. It was then tested on another independent database (DIARETDB1) with the overall average sensitivity, specificity and accuracy of 100%, 96.2% and 97.8%, respectively. In summary, the evaluation results on both retinal image databases demonstrated the effectiveness of using GANs for automatic HE detection.