研究目的
To propose a low-rank approximation based approach for denoising of low-dose CT images by effectively utilizing global spatial correlation and local smoothness properties, integrating tensor nuclear norm and tensor total variation regularization into a unified framework.
研究成果
The proposed method effectively denoises low-dose CT images by integrating tensor nuclear norm and tensor total variation, preserving edges and reducing noise. Experimental results on simulated and real data show superior performance over existing techniques in terms of PSNR, EPI, FSIM, RMSE, and entropy. The ADMM-based algorithm is efficient, and the approach is applicable to both Gaussian and mixed Poisson-Gaussian noise with pre-processing.
研究不足
The method assumes additive Gaussian noise in CT images, which may not always hold true as noise can be mixed Poisson-Gaussian. It requires parameter tuning (e.g., patch size, number of similar patches, scaling factor) for optimal performance, and computational complexity is high due to tensor operations. It may not perform well without proper database, but aims to avoid rigorous learning phases.
1：Experimental Design and Method Selection:
The methodology involves low-rank tensor modeling and total variation regularization for denoising CT images. The tensor nuclear norm captures global spatial correlation, and tensor total variation ensures local smoothness and edge preservation. The optimization problem is solved using the Alternative Direction Method of Multipliers (ADMM).
2：Sample Selection and Data Sources:
Simulated CT images (e.g., Shepp-Logan head phantom, thoracic anthropomorphic phantom) and real CT images from databases (e.g., AAPM Low-Dose CT Grand Challenge, http://www.via.cornell.edu/databases) are used. Images are 3D with sizes like 256x256x100 or 512x512x
3：List of Experimental Equipment and Materials:
CT scanners (Somatom Definition AS+ or Somatom Definition Flash, Siemens Healthcare), software (Matlab 2017A), computer (Intel processor 3.30 GHz CPU, 16 GB RAM).
4：30 GHz CPU, 16 GB RAM).
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: Extract overlapping 3D patches from noisy CT images, group similar patches based on l2-distance, apply t-SVD for low-rank approximation and tensor TV for regularization using ADMM, reconstruct denoised image by averaging patches.
5：Data Analysis Methods:
Performance metrics include PSNR, EPI, FSIM, RMSE, and entropy. Comparisons are made with state-of-the-art methods like INLM, ANLM, KSVD, BM3D, FCR, and LNLM.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
