摘要
视网膜血管的精确分割是眼部疾病临床诊断的关键步骤,针对眼底视网膜血管结构复杂、形态多样、边缘末端细微导致分割难度高的问题,提出了一种基于UNet++新的眼底视网膜血管分割网络模型(Dil-UNet++),该方法在传统的UNet++网络基础上使用多层空洞卷积模块来实现特征提取,使网络获得大小可变的特征提取感受野,获得了更好的特征提取性能;同时在网络跳跃连接部分增加注意力机制模块提高了分割目标相关的通道权重和空间权重;并使用卷积模块代替最大池化模块实现下采样,来避免特征传递时血管细节特征丢失。在分割DRIVE视网膜血管图像数据集上的训练与验证结果表明,该模型在Dice系数、准确度和精确度分别达到87.65%、96.99%和92.82%,证明了该网络模型的有效性,从而为眼底视网膜血管图像分割提供了新的方法。
The accurate segmentation of retinal blood vessels is a critical step in the clinical diagnosis of ocular disorders. However, the vessel segmentation remains highly challenging due to complex structures, as well as blurred edges and variable morphology. Based on UNet++, we propose a new network model for segmenting retinal vessels (Dil-UNet++). The Dil-UNet++ employs multi-layered void convolutional modules for feature extraction, enabling the network to achieve a flexible feature ex-traction receptive field, thereby enhancing the performance of feature extraction. The addition of the attention mechanism module in the jump connection part of the network improves the channel weight and spatial weight associated with the segmentation target. Instead of maximum pooling, convolution is utilized for down-sampling to avoid loss of vessel details. Based on the segmented DRIVE retinal vessel image dataset, the training and validation results show that the Dice coefficient, ac-curacy, and precision are respectively 87.65%, 96.99%, and 92.82%, which prove the validity of Dil-UNet++ model.
出处
《计算机科学与应用》
2024年第1期54-67,共14页
Computer Science and Application
