先验GAN的CBCT牙齿图像超分辨率方法

CBCT Tooth Images Super-Resolution Method Based on GAN Prior

针对高分辨率的锥形束计算机断层扫描图像难以获取的问题,提出一种基于先验生成对抗网络的锥形束计算机断层扫描图像超分辨率方法,其通过微型断层扫描图像作为参照,对超分辨率网络进行弱监督训练.首先训练一个生成对抗网络,用于生成高质量的单颗牙齿的微型计算机断层扫描图像,并将其嵌入到一个U型的主干网络中作为先验解码器;然后利用低分辨率的多颗牙齿的锥形束计算机断层扫描图像对主干网络进行训练,先定位到锥形束计算机断层扫描影像中的每一颗牙,再分别提高分辨率;最后解决锥形束计算机断层扫描与微型计算机断层扫描之间的域差距,通过设计基于小波变换噪声提取的域适应退化模块,间接优化生成器生成更符合微型计算机断层扫描信息分布的图像.在锥形束计算机断层扫描数据集上进行实验的结果表明,与现有的超分辨率方法相比,所提方法的峰值信噪比提高了0.79~6.02 dB,感知相似度评价指标降低了0.01~0.72,并且在牙齿部分获得了更好的视觉效果,具有较强的竞争力.

Aiming at the problem that high-resolution cone-beam computed tomography images are difficult to obtain,a cone-beam computed tomography image super-resolution method based on prior generative adversarial network is proposed,which uses the micro computed tomography image as a reference to perform weakly supervised training on the super-resolution network.Firstly,a generative adversarial network is trained to generate high-quality micro computed tomography images of a single tooth,which is embedded into a U-shaped backbone network as a prior decoder.Then,low-resolution cone-beam computed tomography images of multiple teeth are used to train the backbone network,and the network is first located to each tooth in the cone-beam computed tomography image.Finally,the domain gap between cone-beam computed tomography and micro computed tomography is solved.By designing a domain adaptive degradation module based on wavelet transform noise extraction,the generator is indirectly optimized to generate images more in line with the information distribution of micro computed tomography.The experimental results on the cone-beam computed tomography dataset show that compared with the existing super-resolution methods,the proposed method improves the peak signal-to-noise ratio by 0.79−6.02 dB,reduces the learned perceptual image patch similarity by 0.01−0.72,and has better visual effect in the tooth super-resolution,which has strong competitiveness.