Group Sparse based Super-resolution of Magnetic Resonance Images for Superior Lesion Diagnosis

In the modern times, retrieval of significant data from low-resolution (LR) magnetic resonance (MR) images has turned out to be a strenuous task. Also, in the recent years, several Super-resolution (SR) techniques have been established to address the issue of MR image resolution. This research focuses on developing a Super-resolution MR Image restoration method using group-based sparse representation technique (GSR). The major objective is to devise a GSR technique which is robust to noise, while most other SR methods cannot perform de-noising and super-resolution simultaneously. Moreover, the restoration dependent approach presumes that the LR images are warped, blurred and decimated from the respective high-resolution (HR) image. The algorithm exploits the similarity between non-locally positioned similar patches to effectively improve the quality of MR images. A single self-adaptive dictionary with low-complexity is used in the model in place of the general dictionary used in traditional approaches. This self-adaptive dictionary is trained for a group of patches rather than for each patch. Training the model for a group instead of patches allows the model to have a better edge and texture retention in the reconstructed image. This approach also establishes the fact that an enhanced detection of lesions is highly possible for superior disease diagnosis. The GSR approach proves to be efficient as it offers better PSNR values for all the MR images than its counterparts.