针对视网膜图像中血管细小而导致其分割精度低等问题,通过在U-Net网络中引入Inception、空洞卷积与注意力机制等模块,提出一种改进U-Net视网膜血管图像的分割算法。首先,在编码阶段增加Inception模块,采用不同尺度的卷积核对图像进行特征提取,以获得其多尺度信息;然后,在U-Net网络的底部增加级联空洞卷积模块,以在不增加网络参数的情况下扩大卷积操作的感受野;最后,在解码阶段为反卷积操作设计了注意力机制,将注意力机制与跳跃连接方式相结合,聚焦目标特征,以解决权重分散等问题。基于标准图像集DRIVE的实验结果表明,所提算法的平均准确率、灵敏度与特异性较之U-Net算法分别提高1.15%,6.15%与0.67%,也优于其他传统分割算法。

In this study, we propose an improved U-Net retinal vascular image segmentation algorithm by introducing some modules, such as inception, hole convolution, and attention mechanism, into the U-Net network to solve the problem of low segmentation accuracy caused by the small blood vessels in the retinal image. Initially, the inception module was added during the encoding stage, and convolution kernels of different scales were used to extract the image features to obtain multiscale information from the image. Subsequently, a cascaded hole convolution module was added to the bottom of the U-Net network for expanding the receptive field of the convolution operation without increasing the network parameters. Finally, an attention mechanism was designed for the deconvolution operation during the decoding phase. The problem of weight dispersion can be solved by focusing on the target features under the combination of the attention mechanism and jump connection mode. The experimental results obtained using the standard image set DRIVE denote that the average accuracy, sensitivity, and specificity of the proposed algorithm are 1.15%, 6.15%, and 0.67% higher than those of the traditional U-Net algorithm, respectively, and that the proposed algorithm outperforms other traditional segmentation algorithms.