Research On Nonlocal Means Filter Based Reconstruction Of Dynamic MRI

Magnetic Resonance Imaging (MRI) is an important imaging modality based on Nuclear Magnetic Resonance (NMR). MRI has been widely used in medical applications due to its irreplaceable advantages compared to other imaging modalities such as X-ray, Computer Tomography (CT) and ultrasound. Especially with the advent of dynamic MRI which has aroused general interest, MRI is able to monitor the variations of the imaged object effectively. Dynamic MRI has played an important role in both medical applications and scientific research.In order to effectively capture an instantaneous snapshot of the moving object, the signal acquisition rate need to be further increased. However the signal acquisition of a single slice in MRI is a time consuming process, let alone multiple slices in dynamic MRI. This results in at least two effects. First the overlong process is adverse to capture dynamic information. Second it makes MRI more susceptible to the artifacts caused by rigid motion of patients. The acquisition process can be accelerated through the improvement of hardware and partial k-space acquisition. Nonetheless, the acceleration through hardware (gradient slew rate) is limited by physical and physiological constraints, which enables partial k-space acquisition to be the most important means to accelerate the acquisition process. Many strategies have been proposed to reconstruct images of favorable temporal resolution and spatial resolution with a small portion of k-space data.In this paper, we first analyzed many state-of-the-art fast imaging methods that were grouped into three categories.1. Parallel imaging methods acquire partial lines in phase encoding direction simultaneously with phased array coils and recover the missing part with coil sensitivity or linear correlation between k-space lines.2.View sharing methods only update a certain part of k-space data according to some prior information about the imaged object and fill the missing data with those from adjacent time frames or pre-scanned ones.3. Reconstruction methods of sparsely sampled data, such as CS, k-t SENSE and HYPR. And then we gave a brief review of the characteristics of sampling in dynamic MRI. Finally we proposed an approach named MNLM (Modified Non-local Means), together with other excellent fast imaging methods such as parallel imaging, non-Cartesian sampling and Gridding algorithm etc. to present good reconstruction results with sparsely sampled data in dynamic MRI.Guided by the concept of making full use of the data consisting of all the frames to assist the reconstruction of each individual time frame, we proposed a coil-sensitivity free, non-iterative method for improved reconstruction of highly-undersampled time-resolved MRI based on modified non-local means (NLM), thus named MNLM. In MNLM, two images are simply reconstructed (sliding-window in time and Gridding algorithm) first:image of high temporal resolution reconstructed from less data and image of high spatial resolution reconstructed from more data. Then non-local means filter is introduced to search similar structures for high temporal resolution image in high spatial resolution image, and quantify the similarity with a weighting factor. Finally these two images are combined with the weighting factors to generate an image with both favorable temporal and spatial resolution. In addition, unlike the traditional non-local means filter (in which the search of similar structures is carried out in the restored image itself), instead of calculating filtering parameters based on the information of one image, in MNLM the calculation of the filtering parameters takes two images into account, which will be detailed in this paper. The experiment results of both clinical data and phantom data show that the proposed algorithm can effectively eliminate noise and the artifacts caused by highly-undersampled pattern, enhancing details for image of high temporal resolution.