CT image sequence restoration based on sparse and low-rank decomposition

• Published in: PloS one Vol. 8; no. 9; p. e72696
• Main Authors: Gou, Shuiping, Wang, Yueyue, Wang, Zhilong, Peng, Yong, Zhang, Xiaopeng, Jiao, Licheng, Wu, Jianshe
• Format: Journal Article
• Language: English
• Published: United States Public Library of Science 04-09-2013; Public Library of Science (PLoS)
• Subjects: Algorithms; Artificial intelligence; Boundaries; CAT scans; Computed tomography; Data integrity; Decomposition; Diagnostic imaging; Education; Filtration; Hospitals; Humans; Image contrast; Image Processing, Computer-Assisted - methods; Image restoration; International conferences; Linear programming; Mathematical models; Medical imaging; Medical imaging equipment; Methods; NMR; Noise; Nuclear magnetic resonance; Numerical experiments; Point spread functions; Principal components analysis; Radiographic Image Enhancement - methods; Rankings; Recovery; Robustness (mathematics); Signal processing; Sparsity; Tomography, X-Ray Computed - methods; Wiener filtering; Beijing China; China
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0072696

Blurry organ boundaries and soft tissue structures present a major challenge in biomedical image restoration. In this paper, we propose a low-rank decomposition-based method for computed tomography (CT) image sequence restoration, where the CT image sequence is decomposed into a sparse component and a low-rank component. A new point spread function of Weiner filter is employed to efficiently remove blur in the sparse component; a wiener filtering with the Gaussian PSF is used to recover the average image of the low-rank component. And then we get the recovered CT image sequence by combining the recovery low-rank image with all recovery sparse image sequence. Our method achieves restoration results with higher contrast, sharper organ boundaries and richer soft tissue structure information, compared with existing CT image restoration methods. The robustness of our method was assessed with numerical experiments using three different low-rank models: Robust Principle Component Analysis (RPCA), Linearized Alternating Direction Method with Adaptive Penalty (LADMAP) and Go Decomposition (GoDec). Experimental results demonstrated that the RPCA model was the most suitable for the small noise CT images whereas the GoDec model was the best for the large noisy CT images.