南方医科大学学报

南方医科大学学报 ›› 2021, Vol. 41 ›› Issue (8): 1250-1259.doi: 10.12122/j.issn.1673-4254.2021.08.18

• • 上一篇    下一篇

眼底图像硬渗出物的分割算法：基于区域分类引导的小波Y-Net的EX分割

张利云，方智文 ，唐宇姣，杨 丰   

  1. 南方医科大学生物医学工程学院//广东省医学图像处理重点实验室，广东 广州 510515
  • 出版日期:2021-08-20 发布日期:2021-09-07

Regional classification-guided wavelet Y-Net network for hard exudate segmentation in fundus images

ZHANG Liyun, FANG Zhiwen, TANG Yujiao, YANG Feng   

  1. School of Biomedical Engineering, Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, Guangzhou 510515, China
  • Online:2021-08-20 Published:2021-09-07

RichHTML

49

PDF

277

摘要： 目的 为消除视盘在硬渗出物(EX)分割过程中带来的影响提出了基于区域分类引导的小波Y-Net网络的EX分割算法。方法 该网络为端到端的眼底图像EX分割网络，通过区域分类引导EX分割联合实现了视盘区域检测和EX分割，有效地降低了视盘对EX分割的干扰。为了避免因下采样操作产生信息损失而导致微小EX区域分割失效的问题，该网络进一步引入了离散小波变换（DWT）和离散小波逆变换（IDWT）取代传统的池化下采样和上采样操作。同时，采用了基于残差连接的Inception模块获取多尺度特征。所提出的算法在IDRiD、e-ophtha EX数据库上进行训练和测试，并进行像素级评估。结果 区域分类引导的小波Y-Net网络在IDRiD、e-ophtha EX数据库上分别获得0.9858、0.9938的准确率以及0.9880、0.9986的受试者工作特征曲线下面积（AUC）。结论 本文提出的方法能够有效地规避视盘的影响，保留图像细节信息，提升EX的分割效果。

关键词: 眼底图像；硬性渗出；分割；Inception模块；离散小波变换；Y-Net

Abstract: Objective We propose an hard exudate (EX) segmentation algorithm based on regional classification-guided wavelet Y-Net network to eliminate the influence of optic disc on EX segmentation process. Methods The wavelet Y-Net network was an end-to-end fundus image EX segmentation network, which combined the regional detection of optic disc and hard exudates segmentation by regional classification-guided EX segmentation to effectively reduce the interference of optic disc in EX segmentation. To avoid failure of small EX region segmentation caused by information loss due to down-sampling operation, discrete wavelet transform (DWT) and inverse discrete wavelet transform (IDWT) were introduced to replace the traditional pooling down-sampling and up-sampling operations. Meanwhile, the inception module based on residual connection was used to obtain the multi-scale features. The proposed algorithm was trained and tested on the IDRiD and e-ophtha EX datasets and evaluated at the pixel level. Results For IDRiD and e-ophtha EX datasets, the proposed algorithm achieved accuracy rates of 0.9858 and 0.9938 with AUC values of 0.9880 and 0.9986, respectively. Conclusion The proposed method can effectively avoid the influence of the optic disc, retain the image details, and improve the effect of EX segmentation.

Key words: fundus images; hard exudates; segmentation; inception model; discrete wavelet transform; Y-Net