基于GAN的肺部CT影像超分辨率重建研究

Research on super-resolution reconstruction of lung CT images based on GAN

基于提高CT图像分辨率、丰富放大后图像纹理细节的目的,提出了RUAGAN模型。通过在生成器部分使用RRDB作为基本块,并加入局部注意力(Local aware Attention,LA)的方法提取影像中的高频信息;为提供更加真实的梯度反馈,使用U-Net网络作为鉴别器。经过实验仿真表明,改进后的模型在肺部CT图像数据集上训练,在4x放大因子上重建出的图像纹理更加丰富,有清晰真实的边缘,其峰值信噪比与结构相似性分别达到31.980和0.974,高于其他分辨率重建模型。

To improve the resolution of CT images and enrich the texture details of enlarged images,a RUAGAN model is proposed.In the generator part,RRDB is used as the basic block,and Local aware Attention(LA)is added to extract the high-frequency information in the image.In order to provide more realistic gradient feedback,the U-Net network is used as the discriminator.The experimental simulation results show that the improved model is trained on the lung CT image dataset.The image texture reconstructed on a 4x magnification factor is more affluent,with transparent and real edges.The peak signal-to-noise ratio and structural similarity of the improved model are 31.980 and 0.974,respectively,which are higher than other resolution reconstruction models.