跨级可变形Transformer编解码视网膜图像分割算法  

Cross-stage Deformable Transformer Encoding and Decoding Algorithm for Retinal Image Segmentation

作　　者：梁礼明 阳渊 何安军 李仁杰 LIANG Liming;YANG Yuan;HE Anjun;LI Renjie(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341411,China)

机构地区：[1]江西理工大学电气工程与自动化学院,江西赣州341411

出　　处：《无线电工程》2023年第9期1990-2001,共12页Radio Engineering

基　　金：国家自然科学基金(51365017,61463018);江西省自然科学基金面上项目(20192BAB205084);江西省教育厅科学技术研究重点项目(GJJ170491);江西省研究生创新专项资金项目(YC2022-S676)。

摘　　要：眼底视网膜血管图像分割对青光眼、糖尿病等疾病的预防和诊断具有重要意义。针对视网膜血管图像边缘分割模糊、微细血管漏缺和模型感受野不足等问题,提出了一种跨级可变形Transformer编解码(Cross-stage Deformable Transformer Encoding and Decoding Net,CTED-Net)视网膜图像分割算法。在特征提取网络中融入通道像素增强模块和跨级融合骨干,实现对视网膜血管全局特征的粗提取;在网络编码部分加入可变形自适应编码Transformer模块(Deformable Adaptive Coding Transformer Module,DACT),通过可变形编码的方式增大模型感受野;在编解码结构底层加入深层语义门控注意模块,实现对视网膜血管深层特征的充分学习,以改善血管图像边缘分割模糊的问题。在模型训练阶段采用加权交叉焦点损失函数,克服视网膜血管图像样本不平衡的问题。在公共数据集DRIVE和STARE上进行仿真实验,所提算法灵敏度、特异性、准确率和AUC指标在DRIVE上达到84.25%、98.17%、96.46%和98.70%,在STARE上达到80.22%、98.64%、96.71%和98.78%。通过与其他先进算法对比分析可以看出,所提算法分割效果可靠且整体性能先进。Eyeground retinal vascular image segmentation is of great significance for the prevention and diagnosis of glaucoma,diabetes and other diseases.To solve the problems of blurred retinal vessel edge segmentation,micro-vessel leakage and model receptive field deficiency,a Cross-stage Deformable Transformer Encoding and Decoding Net(CTED-Net)algorithm for retinal segmentation is proposed.Firstly,a channel pixel enhancement module and a cross-level fusion backbone are integrated into the feature extraction network to achieve coarse extraction of global features of retinal vessels.Then,a deformable adaptive coding Transformer module is added to the network coding part,which increases the receptive field of the model through deformable coding.Finally,a deep semantic gated attention module is added to the bottom of the encoding and decoding structure to realize the full learning of the deep semantic of the retina,so as to solve the fuzzy segmentation problem of vascular edge.In addition,the weighted cross-focus loss function is used in the model training stage to solve the imbalance problem of retinal blood vessel image samples.The sensitivity,specificity,accuracy and AUC of the proposed algorithm are 84.25%,98.17%,96.46%and 98.70%on DRIVE and 80.22%,98.64%,96.71%and 98.78%on STARE.Compared with other advanced algorithms,the segmentation effect of the proposed algorithm is reliable and the overall performance is advanced.

关 键 词：可变形Transformer 跨级融合骨干 加权交叉焦点损失函数 视网膜血管图像分割 深层语义门控注意模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]