为了将超分辨率重建算法应用于医学影像领域, 提升各类医学影像的分辨率, 针对当前主流算法网络结构和分辨率提升倍数的尺度单一性问题, 提出了一种应用于CT图像的多尺度残差网络模型。首先, 通过级联多层残差块构建模型框架, 残差块内采用3种尺度的卷积核提取低分辨率图像的细节特征。然后, 将特征图融合在一个维度进行特征映射和数据降维, 并将多尺度特征信息导入下一残差块。最后, 将网络学习到的残差图与低分辨率图像融合, 重建高分辨率图像。采用经过多种放大倍数处理的CT图像对网络进行混合训练, 实现了一个模型可以同时支持多种倍数的分辨率提升。实验结果表明: 在2, 3, 4倍放大因子下, 该模型重建的CT图像PSNR平均较VDSR算法高0.87, 083, 1.16 dB。因此, 本文模型有效提升了CT图像的超分辨率重建效果, 更锐利地恢复了其细节特征, 同时大大提升了算法实用性。

In order to apply the super-resolution reconstruction algorithm to the field of medical imaging and improve the resolution of various medical images, a multi-scale residual network model applied to CT images is proposed for the problem of the singleness of the network structure and the resolution multiplier of the current mainstream algorithm. Firstly, the model framework is built by cascading multi-level residual blocks, and the convolution kernels of three scales are used in the residual block to extract the detailed features of low-resolution images. Then, the feature map is fused in one dimension for feature mapping and data dimensionality reduction, and the multi-scale feature information is imported into the next residual block. Finally, the residual map calculated by the network is merged with the low-resolution image to reconstruct the high-resolution image. The network is trained by CT images processed by multiple magnifications, and so that a model can support multiple resolution enhancements at the same time. The experimental results show that under the 2, 3, and 4 times magnification factors, the PSNR of the reconstructed CT image is 0.87, 0.83, 1.16 dB higher than the VDSR algorithm. Therefore, the model of this paper effectively improves the super-resolution reconstruction effect of CT images, restores its detailed features more sharply, and greatly improves the practicality of the algorithm.