基于特征图金字塔的冠脉造影图像血管分割方法

Coronary artery angiography image vessel segmentation method based on feature pyramid network

针对冠脉造影图像照明不均、血管结构与背景区域对比度低、冠脉血管拓扑结构复杂等分割难点,建立了一个冠脉造影血管分割标注数据集,并在此基础上提出了一种基于特征图金字塔的冠脉造影图像血管分割模型。本文模型以U-Net网络为基础进行改进和优化,首先,将U-Net编码部分的第一个卷积层修改为一个7×7的卷积层,并提高每一层的感受野,在编解码层中引入修改后的ConvNeXt block,使得网络提取更深层次特征的能力有所提升;其次,设计分组注意力机制模块GA,并将其引入到U-Net跨连接处,对编码部分提取的特征进行增强,弥补编解码器间存在的语义差距;最后,在U-Net解码器处设计了一个特征图金字塔级联模块PFC,融合各尺度的特征图,并在PFC中每一层中加入SE注意力机制模块,用于筛选特征图中的有效信息,网络损失函数为PFC模块各层输出的加权,以监督网络各层的特征提取。本文模型在测试集上的测试结果如下:Dice系数为0.8843,Jaccard系数为0.7926。实验结果表明,相比其他常用方法,本文模型在冠脉血管分割上具有较强的鲁棒性,在低对比度下能够有效抑制噪声,对冠脉血管具有更好的分割效果。

To address issues such as uneven illumination in coronary angiography images,low contrast between vascular structures and background regions,and the complexity of coronary vascular topology,we establish a coronary angiography vascular segmentation annotation dataset.Additionally,we propose a coronary angiography image vascular segmentation model based on the feature map pyramid.On the basis of the U-Net architecture,this model was improved and optimized.First,the first convolutional layer in the U-Net encoding part was replaced with a 7×7 convolutional layer to increase the receptive field of each layer.Modified ConvNeXt blocks were added to the encoding and decoding layers to enhance the network's ability to extract deeper-level features.Second,a Group Attention(GA)mechanism module was designed and incor-porated at the U-Net skip connection to strengthen the features extracted from the encoding part,addressing semantic gaps between the encoder and decoder.Finally,a Pyramid Feature Concatenation(PFC)module was designed at the U-Net decoder,which fused features from different scales.Squeeze-and-Excitaton(SE)attention mechanisms were added to each layer of the PFC to filter out effective information from the feature maps.The loss function of the network is weighted based on the outputs of the PFC module at each layer,serving to supervise the feature extraction process across different layers of the network.The test results of this model on the test set are as follows:the Dice coefficient is 0.8843 and the Jaccard coefficient is 0.7926.Experimental results indicate that this model is highly robust in coronary vascular segmentation,more effect-ively suppressing noise under low contrast and achieving better segmentation results for coronary vessels when compared to other methods.