Low-dose CT denoising based on multi-scale dynamic convolution and edge enhancement
Computed Tomography(CT)technology is widely used for disease detection and screening.How-ever,the X-ray radiation generated during the scanning process can harm to human health.Employing low-dose CT reduces radiation exposure to patients,but the images reconstructed at lower doses are severely de-graded 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 re-gions,often leading to over-smoothed results.To migrate this issue,this paper presents MDCEENet,a net-work that incorporates multi-scale dynamic convolution and edge enhancement for low-dose CT image denois-ing,which preserves more image textures and structural details throughout the denoising process.MD-CEENet 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 Informa-tion 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 de-coder.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 avail-able 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 aver-age Peak Signal-to-Noise Ratio(PSNR)and Structural Similarity Index(SSIM),which highlighted the su-periority 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.
Deep learningLow-dose CT denoisingMulti-scale dynamic convolutionEdge enhancement
国家科技期刊平台
NSTL
万方数据
