摘要
生物医学图像分割已成为医学诊断中的关键任务之一。然而,由于组织和器官的复杂形态及其结构的多样性,医学图像分割技术的实际应用面临显著的技术挑战。在传统卷积神经网络(Convolutional Neural Network,CNN)中,最大池化操作常常导致信息的不可逆丢失,尽管引入小波变换在一定程度上改善了这一问题,但小波变换本身也存在局限性。为了解决这一问题,提出了一种基于双树复小波变换(Dual-Tree Complex Wavelet Transform,DTCWT)和U-Net的视网膜血管分割模型——DTCWU-Net。该模型通过引入DTCWT替代传统池化层,双树复小波逆变换(Inverse DTCWT,IDTCWT)替代传统上采样层,显著增强了特征提取能力,尤其在保留图像细节方面表现出色。DTCWU-Net还引入了高低频特征融合注意力(Low and High Feature Fusion Attention,LHFFA)和多尺度门控注意力(Multi-Scale Gate Attention,MSGA)模块,进一步提升分割性能。实验结果表明,DTCWU-Net在DRIVE数据集上取得的准确率(Accuracy,ACC)为0.9686,受试者工作特征(Receiver Operating Characteristic,ROC)曲线下面积(Area Under the ROC Curve,AUC)为0.9867,在CHASE_DB1数据集上取得的ACC为0.9750,AUC为0.9903,在STARE数据集上取得的ACC为0.9757,AUC为0.9901。在F1、灵敏度(Sensitivity,SE)、ACC和AUC等关键指标上,超越了其他主流方法的表现。通过多模块协同优化,DTCWU-Net显著提高并展现了视网膜血管分割精度与细节恢复能力。
Biomedical image segmentation has become one of the key tasks in medical diagnosis.However,due to the complex morphology of tissues and organs and the diversity of their structures,the practical application of medical image segmentation techniques faces significant technical challenges.In traditional Convolutional Neural Network(CNN),the maximum pooling operation often results in irreversible loss of information.Although the incorporation of the wavelet transform mitigates this issue to some extent,the wavelet transform itself also has its own limitations.To solve this problem,a retinal blood vessel segmentation model named DTCWU-Net based on Dual-Tree Complex Wavelet Transform(DTCWT)and U-Net is proposed.The model replaces the traditional pooling layer with DTCWT and traditional upsampling layer with the Inverse DTCWT(IDTCWT).This significantly enhances the feature extraction ability,particularly in preserving image details.DTCWU-Net also introduces Low and High Feature Fusion Attention(LHFFA)and Multi-Scale Gate Attention(MSGA)modules to further improve the segmentation performance.The experimental results show that DTCWU-Net achieved an Accuracy(ACC)of 0.9686 and an Area Under the ROC(Receiver Operating Characteristic)Curve(AUC)of 0.9867 on the DRIVE dataset,an ACC of 0.9750 and an AUC of 0.9903 on the CHASE_DB1 dataset,an ACC of 0.9757 and an AUC of 0.9901 on the STARE dataset,The proposed model demonstrates superior performance in key metrics when compared to other mainstream methods,including F1,Sensitivity(SE),ACC and AUC.Through the collaborative optimization of multiple modules,DTCWU-Net significantly improves and demonstrates the accuracy of retinal vessel segmentation and the ability to recover details.
作者
陶寅涵
朱家明
吴军
TAO Yinhan;ZHU Jiaming;WU Jun(College of Information Engineering,Yangzhou University,Yangzhou 225127,China)
出处
《无线电工程》
2025年第6期1161-1176,共16页
Radio Engineering
基金
国家自然科学基金(62473362)。
