基于注意力门UNet网络的CT金属伪影去除方法

Metal Artifact Reduction for CT Method Based on Attention Gate UNet

摘要:

目前,UNet基本模型对带有金属伪影的CT图像的去除能力无法有效满足需求,UNet的结构简单无法提取出足够精确的有效结构和细节信息,并且深层卷积对低级特征的信息利用不够充分。基于上述问题,提出了一个基于注意力门的UNet金属伪影去除网络,该网络采用了注意力门对低层级和高层级的信息进行注意力权重处理,并利用跳跃连接机制到特征解码结构以提高生成CT图像的质量,通过多层级的编解码结构得到最终的去除金属伪影CT图像。实验结果表明,该方法与ADN、cGANMAR、UNet、CNNMAR、CycleGAN相比得到了更好的条状和带状伪影去除效果的CT图像。

关键词:金属伪影去除;注意力机制;卷积神经网络;UNet;编解码

Abstract:

Currently, the basic UNet model cannot effectively meet the demand for CT images with metal artifacts reduction, the simple structure of UNet cannot extract sufficiently accurate information about the effective structure and details, and the deep convolution does not sufficiently use the information of low-level features. Based on the above problems, a metal artifact training dataset based on actual CT images and a metal artifact removal network with attention gates based on UNet is constructed. The network adopts attention gates to apply attention weights to the information at low and high levels and feeds back to the feature decoding structure using a jump connection mechanism to improve the quality of the generated CT images. The final CT image Experience results show that the method yields CT images with better removal of stripe and band artifacts compared to ADN,cGANMAR,UNet,CNNMAR,CycleGAN models.

Key words:digital watermark; wavelet packet; adaptive; feature of texture