融合感受野模块的卷积神经网络视杯视盘联合分割

青光眼是世界第一大不可逆致盲性眼病,早期诊断和及时治疗是预防青光眼致盲的有效措施。眼底图像中的杯盘比是青光眼早期筛查与临床诊断的重要指标。因此,精确的视杯视盘分割是准确计算杯盘比并提高青光眼计算机辅助诊断技术的关键。针对这一问题,在对眼底图像进行极坐标变换的基础上,提出一种融合感受野模块的卷积神经网络Seg-RFNet,以实现视杯视盘联合分割。Seg-RFNet以SegNet框架为基础,使用ResNet50作为编码层,增强图像的特征提取能力,并对编码层进行分支处理,进一步获得更多的深层语义信息;同时在编码层和解码层之间加入感受野模块,模拟人类视觉系统,在增大感受野的同时增强了有用特征的响应。使用MICCAI 2018公开数据集REFUGE中的800张眼底图像对所提出方法与其他方法进行性能验证和比较。结果表明,Seg-RFNet分割视杯和视盘的Jaccard相似度分别0.951 5和0.872 0,F分数达到了0.974 9和0.930 1,与常用的U-Net、SegNet 网络相比,Seg-RFNet具有更好的视杯视盘联合分割精度,为计算杯盘比提供了精确分割基础。

Joint Optic Cup and Disc Segmentation Using Convolutional Neural Network with Receptive Field Module

Glaucoma is the world's largest irreversible blindness eye disease. Early diagnosis and timely treatment are effective measures to prevent blindness caused by glaucoma. The cup to disc ratio in fundus images is an important index of early screening and clinical diagnosis of glaucoma. Therefore, accurate segmentation of the optic cup and disc is the key to accurately calculate the cup to disc ratio and improve the computer-aided diagnosis technology of glaucoma. To solve this problem, this paper firstly performed a polar coordinate transformation preprocessing on the fundus image, and then proposed a convolutional neural network Seg-RFNet that integrated the receptive field module to achieve joint segmentation of the optic cup and disc. Seg-RFNet was based on SegNet framework and used ResNet50 as the coding layer to enhance the feature extraction ability of the image, and the coding layer was branched to obtain more deep semantic information. At the same time, the receptive field module was added between the coding layer and decoding layer, which was able to simulate the human visual system, increasing the receptive field and enhance the response of useful features. The 800 fundus images from REFUGE that is a dataset published in MICCAI 2018, were used to verify the performance of the proposed method compared with other methods. The results showed that the Jaccard Similarity (higher is better) of the optic cup and disk segmentation was 0.9515 and 0.8720, and the F score (higher is better) was 0.9749 and 0.9301, respectively. Compared with the commonly used U-Net, SegNet and other networks, Seg-RFNet showed better joint segmentation accuracy of the optic cup and disc and provided an accurate segmentation basis for calculating the cup to disc ratio.