Non Local Methods For Denoising Magnetic Resonance Images

Magnetic resonance imaging (MRI) is a medical imaging modality commonly used by radiologists to view living organs and tissues. It reveals important diagnostic information, which may be central to further computer-aided analysis. However, MR images are inherently affected by noise, which degrades the image quality and results in loss of important diagnostic information. Removing this noise using post processing denoising techniques is therefore highly demanding.The objective of an effective denoising algorithm is to restore higher quality images as well as preserve fine details in images. In this thesis, new methods have been proposed for robust denoising by heuristically modifying and combining contemporary denoising methods. The presented methods have been validated on various types of brain MRI sequences. The research work carried out in this thesis has been divided into four phases.In the initial phase, an improved adaptive non-local means (IANLM) algorithm has been proposed as a variation of classical non-local means (NLM) algorithm. The proposed IANLM algorithm adapts the search window size employed in classical NLM with the help of a robust threshold criterion and window adaptation test. Furthermore, IANLM is implemented using an alternate traversal mechanism of search window, which exploits the image homogeneity/smoothness property in order to yield improved denoising performance. Due to adaptive window size, IANLM is also computationally more efficient compared to the classical NLM algorithm. In the second phase, an enhanced NLM (ENLM) algorithm is presented for application to brain MR images, which reveals the real potential of IANLM by adapting it to Rician noise. A bias correction method is employed in ENLM for this purpose, which removes the bias introduced in MR images due to Rician noise. Moreover, a selective median filtering is introduced in ENLM as post processing for improving the quality of homogenous regions of restored images.The selective median filtering procedure is similar to conventional median filtering, but it is only applied to homogeneous regions of images with the help of fuzzy c-means clustering. Finally, empirically suitable values of various parameters of ENLM are obtained for brain MRI denoising.In the third phase, the output of second phase is further refined and an extended NLM (XNLM) algorithm is presented, which adapts IANLM to Rician noise using an improved Rician bias correction method.Two more modifications are presented in XNLM to improve the computational and denoising performance of the algorithm. First, a pixel pre-selection method is proposed to select a subset of pixels within the search window. These pixels are more suitable for computation of restored value of the pixel under consideration. The process pixel pre-selection improves the computational efficiency of XNLM. Moreover, a wavelet coefficients mixing procedure is employed in XNLM in order to exploit capabilities of wavelets for better denoising. Lastly, empirical suitable values of various parameters of XNLM are obtained for brain MRI denoising.The final phase discusses the application perspective of the work done in this research. It investigates the impact of proposed denoising methods on the process of image segmentation, which is usually a subsequent step in a typical medical imaging workflow. It has been shown that the proposed denoising methods improve the performance of segmentation compared to other state of the art segmentation techniques.The denoising methods proposed in this dissertation have been validated on various types of brain MRI sequences including T1-weighted, T2-weighted and PD-weighted MRI. To check the robustness of the proposed methods, images corrupted with various levels of noise have been used in experiments. The superior performance of the proposed methods has been verified from their quantitative and qualitative performance comparison with various state of the art denoising techniques.Hence, the proposed methods are valuable contribution in the field of denoising magnetic resonance images