Atlas-based reconstruction of high performance brain MR data. (April 2018)

• Record Type: Journal Article
• Authors: Zhang, Mingli; Desrosiers, Christian; Zhang, Caiming
• Is Part Of: Pattern recognition. Volume 76(2018:Apr.)
• Page Start: 549
• Page End: 559
• Publication Date: 2018-04
• Subjects: Multi-subject MRI -- Weighted TV -- Sparse representation -- ADMM; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.11.025
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: To our knowledge, this is the first approach to combine internal and external priors in a single consistent model. Internal information is considered as groups of similar patches in the image, which are reconstructed together using multiple sparse dictionaries. These dictionaries are learned with a Gaussian Mixture Model (GMM), providing a more efficient and compact representation of patches. External information is also incorporated in the model in the form of a weighted TV regularization prior, the weights of which are driven by a probabilistic atlas of gradients. These internal and external image priors offer complementary information, the first one modeling nonlocal repetitive patterns and the other one preserving the contours and textures of anatomical structures. An extensive set of experiments is presented for validating the proposed approach. These experiments compare our approach against eight different CS methods on the task of reconstructing brain MR images from undersampled k-space measurements. Results show our approach to outperform state-of-the-art methods for this task. Abstract: Image priors based on total variation (TV) and nonlocal patch similarity have shown to be powerful techniques for the reconstruction of magnetic resonance (MR) images from undersampled k-space measurements. However, due to the uniform regularization of gradients, standard TV approaches often over-smooth edges in the image, resulting in the loss of important details. This paper proposes a novel compressed sensing method which combines both external and internal information for the high-performance reconstruction of MRI data. A probabilistic atlas is used to model the spatial distribution of gradients that correspond to various anatomical structures in the image. This atlas is then employed to control the level of gradient regularization at each image location, within a weighted TV regularization prior. The proposed method also leverages the redundancy of nonlocal similar patches through a sparse representation model. Experiments on T1-weighted images from the ABIDE dataset show the proposed method to outperform state-of-the-art approaches, for different sampling rates and noise levels.