Low-dose CT iterative reconstruction based on image block classification and dictionary learning

For conventional image reconstruction based on dictionary learning in low-dose computed tomography (CT) imaging, all image blocks are represented by the same dictionary, thus limiting the reconstructed image quality. To improve the outcome, a low-dose CT iterative reconstruction algorithm based on image block classification and dictionary learning is proposed. First, each image block is classified as a smooth block or a detail block according to the local image variance. The detail block is subsequently divided into irregular blocks and edge blocks with different angles according to the pointing angle obtained from its gradient field information. Then, the conventional k -singular value decomposition algorithm is applied to train dictionaries for different types of image blocks, and orthogonal matching pursuit determines the sparse coefficients during training. Further, a variety of dictionary learning algorithms are used in penalty-weighted least-squares reconstruction as regular terms. Finally, the relaxed linearized augmented Lagrangian method with ordered subsets is used to solve the objective function. Experimental results show that the proposed algorithm suppresses noise and sharpens edges in reconstructed CT images. The code of the proposed algorithm is available at https://github.com/LIUyi827728/PWLS_BCDL .