【摘要】 双能计算机断层扫描可提供扫描对象的定量信息，实现材料分解，进而获取人体组织材料的特定信息；针对传统的U-Net网络从双能CT图像中提取非局部特征受限的问题，提出了一种改进的U-Net网络，旨在提高双能CT图像材料分解的准确性；IU-Net采用多尺度编码器，通过三条路径从不同角度捕捉输入图像的局部和非局部特征，并在通道维度上进行融合；为避免过度平滑处理导致图像细节丢失，引入了边缘损失，以构建混合损失函数，优化重建图像的边缘像素，产生更清晰的图像；实验结果表明，提出的IU-Net能够保留更多的图像内部细节，分解后的图像更加清晰。更多还原

【Abstract】 Dual-energy computed tomography can provide quantitative information on the scanned object, realize material decomposition, and then obtain specific information of human tissue materials. To solve the problem that the traditional U-Net network is limited in extracting non-local features from dual-energy CT images, an improved U-Net network is proposed to improve the accuracy of dual-energy CT image material decomposition. The inception u-net(IU-Net) uses multi-scale encoders to capture local and non-local features of input images from different angles through three paths, and to fuse them in channel dimensions. In order to avoid excessive smoothing and loss of image details, the edge loss is introduced to construct a hybrid loss function, optimize the edge pixels of the reconstructed image, and produce clearer images. Experimental results show that the proposed IU-Net can retain more internal details of the image, thus making the decomposed image clearer.更多还原