摘要
在眼底图像分析中,视盘分割对于糖尿病视网膜病变、青光眼等眼部疾病的计算机辅助诊断具有重要意义.然而,由于眼底图像数据集存在样本容量有限、数据正负样本不均衡及视盘边缘受血管遮挡严重等问题,这给视盘分割带来了极大挑战.针对以上问题,本文提出一种融合迁移学习和特征金字塔网络的视盘分割模型TriNet,通过迁移学习缓解数据集不足带来的影响;通过结合特征金字塔网络,利用多尺度特征图,提高对视盘的识别率;通过使用Focal Loss损失函数,动态改变正负样本权重,在正负样本失衡的情况下提高网络对稀少样本的识别.在Baidu Research Open-Access Dataset中的iChallenge-AMD项目的眼底图像数据集上进行的仿真结果表明,视盘分割的IoU和DICE精度分别达到了93.48%和96.59%.
In fundus image analysis,accurate segmentation of optic disc has a significant impact on computer-aided diagnosis of diabetic retinopathy,glaucoma and other ocular diseases.Aiming at the difficulties in utilizing the existing algorithms such as the lack of?large-scale datasets,the imbalance of samples and the severe occlusion of the optic disc edge,a optic disc segmentation model TriNet based on transfer learning and feature pyramid networks is proposed,which uses transfer learning to alleviate the impact of insufficient datasets and utilizes the feature pyramid network to obtain multi-scale feature maps so as to improve the detection of the optic disc.The Focal Loss function is used to dynamically change the weight of positive and negative samples,which improves the recognition of sparse samples by the network when positive and negative samples are out of balance.The simulation experiments on iChallenge-AMD fundus image datasets in Baidu Research Open-Access Dataset show that the IoU and DICE precision of our method on optic disc segmentation are 93.48% and 96.59% respectively.
作者
贾西平
黄静琪
罗斌钊
张倩
陈晓静
JIA Xi-ping;HUANG Jing-qi;LUO Bin-zhao;ZHANG Qian;CHEN Xiao-jing(Guangdong Polytechnic Normal University,Guangzhou Guangdong 510665)
出处
《广东技术师范大学学报》
2021年第3期1-7,共7页
Journal of Guangdong Polytechnic Normal University
基金
国家自然科学基金(61872096)
广东普通高校重点项目(2019KZDXM063)
广东省教育厅青年创新人才项目(2016KQNCX092)。
作者简介
贾西平,广东技术师范大学副教授;通讯作者:张倩,博士,广东技术师范大学讲师,E-mail:234482377@qq.com。
