基于域适应对抗网络的眼底图像联合分割方法

Method for Joint Segmentation of Fundus Images Based on Domain Adaptation Adversarial Network

由于数据集之间存在域偏移问题,基于深度学习的语义分割网络在不同数据集之间进行视盘视杯分割性能存在很大差异,这使得不同的医学站点之间进行精确的图像分析和诊断具有一定的挑战性。针对这一问题,提出了一种U-Net结合域对抗网络(domain adversarial via U-Net network,DAUNet)的无监督域适应视盘视杯联合分割方法,并在视盘视杯分割领域取得了不错的性能。首先,利用对抗思想结合目标数据先验特征信息生成与目标数据集相似的数据,预先调整网络参数;其次,通过对抗学习源域和目标域的域变特征,从而降低域偏移的影响,提高分割性能。在REFUGE、Drishti-GS和RIM-ONE-r3共3个数据集之间进行跨数据集的域适应实验和消融实验。实验结果表明,DAUNet网络在以REFUGE作为源域,RIM-ONE-r3作为目标域上视杯的Dice系数,视盘的Dice系数和CDR的绝对错误率分别为0.6486、0.7898、0.0725,优于CADA的分割结果。在消融实验中,视盘分割和视杯分割在有对抗下分别优于无对抗8.00%、4.59%。提出的U型域对抗网络综合了U-Net和域对抗网络(domain-adversarial neural network,DANN)模型的优点,DANN模型中的生成器和判别器联合工作时,会相互对抗并优化分割和判别能力,从而显著提高不同数据集之间的分割性能。

Due to the domain shift problem between datasets,semantic segmentation networks based on deep learning exhibit significant performance variances in optic disc and cup segmentation across different datasets,which makes it challenging to perform accurate image analysis and diagnosis among different medical sites.To address this problem,this paper proposes an unsupervised domain-adapted method for the joint segmentation of the optic disc and optic cup,using a U-Net combined with a Domain Adversarial via U-Net Network(DAUNet),which has achieved excellent performance in the field of optic disc and optic cup.Firstly,by leveraging adversarial principles along with prior feature information of the target data,the method generates data similar to the target dataset,and the network parameters are adjusted in advance;secondly,by adversarial learning of the domain-invariant features of the source and target domains,the influence of domain shift is reduced and the segmentation performance is improved.Cross-dataset domain adaptation experiments and ablation experiments were conducted between three datasets:REFUGE,Drishti-GS and RIM-ONE-r3.The experimental results show that the DAUNet network achieves Dice coefficients of 0.6486,0.7898 and an absolute error rate of CDR of 0.0725 for the cup,disc and CDR respectively on RIM-ONE-r3 as the target domain and REFUGE as the source domain,which are better than those of CADA s segmentation results.In the ablation experiment,disc and cup segmentation were improved by 8.00%and 4.59%,respectively,when utilizing adversarial techniques compared to non-adversarial approaches.The proposed U-shaped domain adversarial network combines the advantages of the U-Net and the Domain-Adversarial Neural Network(DANN)models.When the generator and discriminator in the DANN model work together,they will compete and optimize both the segmentation and discrimination abilities,thus significantly improving the segmentation performance between different datasets.