基于多尺度动态卷积和边缘增强的低剂量CT去噪

Low-dose CT denoising based on multi-scale dynamic convolution and edge enhancement

计算机断层扫描(CT)技术广泛应用于疾病检测与筛查,然而在扫描过程中产生的X射线辐射会对人体造成伤害.采用低剂量CT可以减少患者的辐射暴露,但是重建的图像会有显著的噪声和伪影,干扰医生的诊断.针对这一挑战,众多学者提出了基于传统卷积神经网络的低剂量CT去噪算法,并已取得显著成就.然而,传统卷积在不同像素位置共用相同卷积滤波器,这会忽略不同图像区域的内容差异,导致去噪结果的过度平滑化.为避免这一问题,本文提出一种基于多尺度动态卷积和边缘增强的低剂量CT去噪网络MDCEENet,旨在在去噪过程中保留更多的图像纹理和结构细节.MDCEENet是自编码器结构,包含编码器和解码器两个主要模块.具体而言,将低剂量CT图像及其边缘信息输入到编码器中,通过多尺度特征流MFS和边缘信息流EIS,分别提取多尺度图像特征和图像边缘特征,并将它们融合成引导信息GI,引导解码器中多尺度动态卷积块MDConvBlock的参数生成.在GI的引导下,MDConvBlock模块对上采样特征进行多尺度空洞卷积计算,旨在获取更高质量的重建图像.本文在Mayo Clinic公开的两个数据集上执行了相关实验,通过实验结果可知,MDCEENet的去噪效果优于DnCNN、RED-CNN、WGAN、CNCL、NBNet,获得了最优的平均峰值信噪比和平均结构相似性指标,这表明本文提出方法的优越性.本文还在这两个数据集上进行了消融实验,来说明MDCEENet中引入多尺度动态卷积和边缘信息的有效性,以及与ADFNet网络的区别.实验结果表明了本文提出的方法相比于ADFNet更适用于低剂量CT去噪任务.

Computed Tomography(CT)technology is widely used for disease detection and screening.However,the X-ray radiation generated during the scanning process can harm to human health.Employing lowdose CT reduces radiation exposure to patients,but the images reconstructed at lower doses are severely degraded by noise and artifacts,interfering with a physician's diagnostic process.To address this challenge,many researchers have proposed low-dose CT denoising algorithms based on traditional convolutional neural networks(CNNs),achieving noteworthy successes.However,the traditional convolution applies the same filters uniformly across all pixel positions,which neglects the distinct content features of various image regions,often leading to over-smoothed results.To migrate this issue,this paper presents MDCEENet,a network that incorporates multi-scale dynamic convolution and edge enhancement for low-dose CT image denoising,which preserves more image textures and structural details throughout the denoising process.MDCEENet adopts an autoencoder architecture consisting of an encoder and a decoder.Specifically,low-dose CT images and their corresponding edge information are fed into the encoder,in which multi-scale features and edge features are separately extracted by the Multi-scale Feature Stream(MFS)and the Edge Information Stream(EIS).These features are then integrated into Guidance Information(GI),which in turn guides the parameter generation for the Multi-Scale Dynamic Convolution Block(MDConvBlock)within the decoder.With the guidance of GI,the MDConvBlock applies multi-scale dilated convolutions on the upsampled features to achieve a higher quality image reconstruction.Experiments were conducted on two publicly available datasets from the Mayo Clinic and the experimental results showed that MDCEENet outperforms DnCNN,RED-CNN,WGAN,CNCL,and NBNet in denoising performance,achieving the highest average Peak Signal-to-Noise Ratio(PSNR)and Structural Similarity Index(SSIM),which highlighted the superiority of the proposed method.Furthermore,ablation studies on these datasets validate the effectiveness of incorporating multi-scale dynamic convolution and edge information in MDCEENet,as well as its difference from the ADFNet network.The results indicated that the proposed approach is more suitable for low-dose CT denoising tasks than ADFNet.