Yuanke Zhang1,2, Hongbing Lu1, Junyan Rong1, Yuxiang Xing3, Jing Meng2 Retrieving Adaptive Prior Features from Full-dose Training Database for Low-dose CT Image Restoration Yuanke Zhang1,2, Hongbing Lu1, Junyan Rong1, Yuxiang Xing3, Jing Meng2 1. School of Biomedical Engineering, Fourth Military Medical University, China. 2. School of Information Science and Engineering, Qufu Normal University, China. 3. Department of Engineering Physics, Tsinghua University, China. Introduction The rich structure information in the high quality full-dose CT (FdCT) scans can be exploited as prior knowledge for low-dose CT (LdCT) imaging. For a specific region in an LdCT image, its local structure and texture pattern may be quite different from others. Therefore, locally adaptive prior features would be of great help for the preservation of details/textures in the LdCT image. This study aims to explore a novel prior knowledge retrieval and representation paradigm for capturing local features adaptively from FdCT samples, for the restoration of LdCT images with detailed textures. The innovation is the construction of an offline training database from FdCT scans and the use of online patch- search scheme integrated with the PCA. Methodology The proposed scheme is illustrated by the flowchart shown in Fig. 1. A. This flowchart contains four major steps: Step 1: Offline training database construction Step 2: Online patch search Given a target patch to be restored, we (1) select the most N similar reference patches from the training database as the training samples. (2) select the most N similar noisy patches from its search window on the LdCT image: Step 3: Local principle features retrieved by PCA This process can be expressed by where Step 4: Target patch decomposition and adaptive coefficient shrinkage Target patch decomposition: Coefficient shrinkage: B. An Alternative If No Similar Patch Found in the Training Database Comparing the distance between the target patch and the average patch with a present threshold is adopted to detect whether there are similar patches in the reference database or not. For the target patches that no similar reference patches be found in the database, we instead use the noisy patches as training samples for PCA. Figure 1. Illustration of the flowchart of the proposed algorithm. Numerical Simulation Full-dose lung scans (Siemens CT scanner at 120 kVp and 100 mAs) from 11 patients were used to evaluate the algorithm. Scan slices from patient #1 to #10 were used to construct the offline training database. One scan from the patient #11 was added noise to simulated the low-dose CT image. Results FdCT LdCT denoised References H. Yu, et al. Acad. Radiol., 16: 363–373, 2009. P. Theriault-Lauzier, et al. Med. Phys., 39: 66-80, 2012 Q. Xu, et al. TMI, 31: 1682-1697, 2012. H. Zhang, et al. TMI, 35: 860-870, 2016. L. Zhang, et al. Pattern Recognition, 43: 1531-1549, 2010. Figure 2. Processing results of the simulated low-dose phantom