基于动态特征图卷积网络的视网膜血管分割方法

Retinal Vessel Segmentation Based on Dynamic Feature Graph Convolutional Network

利用深度学习技术进行视网膜血管分割是临床医生诊断眼底疾病的重要辅助方法。然而,现有方法忽略了感受野在高分辨率图像特征提取阶段的重要性以及深度特征图通道之间的动态拓扑相关性,导致通道间信息表征不足。为此,设计一种新颖的动态特征图卷积网络模型(DF-Net)。该模型采用结构化Dropout卷积块替代U-Net中的原始卷积块,构建一个双扩张卷积块(DDCB),旨在不增加模型参数量的同时扩大感受野,提升模型捕捉视网膜血管全局结构信息的能力。进一步,构建动态特征图卷积模块(DFGCM),将特征通道映射到拓扑空间并在拓扑空间上挖掘通道之间的动态拓扑关系,提高通道信息的利用率,丰富视网膜血管的特征信息。实验结果表明:所提DF-Net能够有效提高对视网膜血管的分割精度,在两个公共合理的高分辨率Fives和HRF数据集上,准确率分别达0.9876和0.9733,灵敏度分别达0.9088和0.8322,均优于现有的先进血管分割模型;在三个广泛使用的低分辨率DRIVE、STARE和CHASE_DB1数据集上,也表现出较为出色的分割效果。这些出色的分割结果表明所提DFNet具有辅助临床医生提高眼底疾病诊断能力的潜力。

Objective Various eye diseases and systemic conditions,such as macular degeneration,glaucoma,diabetes,and hypertension,can lead to visual impairment or permanent blindness.These conditions often result in changes in the retinal vascular structure during their onset and development.Ophthalmologists frequently observe abnormalities in retinal blood vessels through fundus imaging.Nevertheless,manual delimitation of retinal vessels in fundus images based solely on expertise and experience is time-consuming,labor-intensive,and subjective,particularly for small/subtle blood vessels.Accordingly,it is important to develop a rapid,accurate,and automated method for retinal blood vessel segmentation using computer vision technology for ophthalmic clinical research.While numerous deep learning methods for retinal blood vessel segmentation have improved the segmentation performance to a certain extent,most of them do not fully consider the complex blood vessel structural features.There remains room for improvement in extracting multi-scale vessel patterns,utilizing global contextual semantic information,and achieving high precision in segmenting small/tiny blood vessels.Therefore,we develop a novel dynamic feature graph convolutional network(DF-Net).Methods The proposed DF-Net,as an end-to-end retinal blood vessel segmentation network,used U-Net as the backbone and consisted of three main parts:an encoder embedding a double dilated convolution block(DDCB),a dynamic feature graph convolution module(DFGCM),and a decoder,as illustrated in Fig.1.The DDCB expanded the receptive field of the model without increasing the number of model parameters,which could better capture global semantic information and improve the model's ability to obtain global structural information on retinal blood vessels.DFGCM captured topological dependencies across feature channels,aggregated effective feature information among channels,and enriched the feature details of the retinal blood vessels.Retinal fundus images were first fed into an encoder integrating DDCB,where the scale of feature maps was reduced by half,and the number of channels was doubled when experiencing each down-sampling operation.Subsequently,the high-level feature maps generated by the encoder were incorporated into the DFGCM.The integration mapped the feature map channels into the topological space,extracting topological correlations among different channels from the topological maps,and aggregating effective features among channels,thereby improving channel utilization.Furthermore,the high-level features output by the DFGCM were fed into the decoder to reconstruct them to the same size as the input image,yielding segmentation results.In the decoder,the scales increased by half and the number of channels was reduced by half when implementing each upsampling operation on the feature maps.The resulting feature maps were then concatenated with the corresponding level maps from the encoder,followed by convolutional operations on these feature maps to extract rich vessel structural information.Results and Discussions Two publicly available high-resolution datasets,Fives and HRF,and three widely used low-resolution datasets,DRIVE,STARE,and CHASE_DB1,were used to validate the proposed DF-Net.The comparison results(Tables 1 and 2)demonstrate that our DF-Net outperforms the existing state-of-the-art retinal vessel segmentation models on the two high-resolution Fives and HRF datasets.For the Fives dataset,our model achieves an accuracy of 0.9876,sensitivity of 0.9088,specificity of 0.99360,AUC of 0.9950,F1-score of 0.9125,and MCC of 0.9059.For the HRF dataset,the model achieves an accuracy of 0.9733,sensitivity of 0.8322,specificity of 0.9837,AUC of 0.9856,F1-score of 0.8318,and MCC of 0.8202.Specifically,compared to U-Net,our model demonstrates improvements of 0.30%,2.60%,0.28%,2.65%,and 2.83%in accuracy,sensitivity,AUC,F1-score,and MCC on the Fives dataset,respectively.Similarly,compared to G-CASCADE,our model achieves improvements of 0.07%,1.16%,0.18%,1.13%,and 1.07%,respectively,on the same dataset.For the HRF dataset,our model achieves performance gains of 0.53%,1.62%,0.05%,0.32%,1.60%,and 1.40%for all the evaluation indices compared with G-CASCADE.The comparison results(see Tables 3,4,and 5)on the three low-resolution DRIVE,STARE,and CHASE_DB1 datasets reveal that our model is relatively more robust and has better generalization capability than other cutting-edge segmentation methods.On the DRIVE dataset,our model achieved the highest scores in terms of accuracy,specificity,and AUC.However,the F1-score,sensitivity,and MCC were 0.83%,2.12%,and 1.16%lower,respectively,than the corresponding highest scores.In the STARE dataset,the gaps in these three indicators are further narrowed,being only 0.50%,0.59%,and 0.55%lower than the corresponding best scores.Our model achieved the highest sensitivity,AUC,F1-score,and MCC for the CHASE_DB1 dataset.The perfect segmentation performance of the developed DF-Net may be attributed to our DDCB for enlarging the receptive field of the model to extract global structural information of retinal blood vessels,as well as DFGCM for establishing dynamic topology relationship among channels to enrich their local characteristics.In addition,the effectiveness of the components of the proposed DF-Net,including DDCB and DFGCM,was justified on the Fives dataset.Conclusions In the present study,a novel DF-Net for high-resolution fundus images is developed to segment retinal blood vessels.It enlarges the receptive field through the DDCB to capture the global structural information of the retinal blood vessels.Additionally,the feature channels are mapped into the topological space using the constructed DFGCM,in which the dynamic topological correlations across channels are deeply extracted,effectively merging the feature information among the channels and refining their local details.Quantitative and qualitative experiments are conducted on two high-resolution fundus image datasets,HRF and Fives,together with three widely used low-resolution datasets:DRIVE,STARE,and CHASE_DB1.The results indicated that the proposed DF-Net outperforms the current advanced retinal blood vessel segmentation methods in most evaluation indicators and exhibits good robustness and generalization capacity.As illustrated in Figs.4 and 5,even in confusing regions where other methods are prone to errors,the proposed model can still correctly identify small/tiny blood vessels in these areas while maintaining the anatomical structures,which are nearly consistent with the ground truths.Thus,we believe that the proposed method can be easily adapted to address other medical image segmentation challenges with diverse appearances and anatomical structures.