ADAPTIVE WAVELET THRESHOLDING FOR NON- HOMOGENEOUS NOISE REDUCTION IN MR (MAGNETIC RESONANCE) IMAGES

Abstract

Noise suppression in MR (Magnetic Resonance) images is a critical task; conventional signal processing techniques are not always suitable as spatial resolution may lose during noise suppression process. Therefore noise suppression ought to be performed in a manner so as to preserve the actual pattern of the image. Non-homogeneous noise is one of the challenges faced in image processing. This thesis work; specifically focuses on non-homogeneous noise suppression method for MR images. Wavelet Analysis has widely been used for image processing including image de-noising, edge detection and segmentation. The existing wavelet de-noising methods are focused on homogeneous noise removal, using same threshold for entire image. If the image contains different burst of random noise, these conventional methods are not sufficient for effective noise removal. The quality of the post-processed image is further affected if these noise patterns cover hard to find malignant areas, which possibly increases the false alarm for diagnostic imaging. In order to improve the early detection of possible malignant areas, the quality of the post-processed image requires effective de-noising techniques, which can be adapted with the nature of noise burst. The fuzzy rule based wavelet thresholding method has been explored in this research for effective noise removal from an image with an array of complexities. In order to develop a robust system closer to real image with non-homogeneous noise, a complex range of noise patterns have been incorporated in MR images. The initial phase of the dissertation work involves the synthesis of non-homogeneous noise on various MR images. Real MR images without noise burst were used as a benchmark. The de- noised images are compared with their clean counterparts for measuring the effectiveness of the technique. A novel image synthesis process has been developed for analyzing the image de- noising and segmentation. Some of the images contain various sizes of malignant patterns for full scale analysis of image de-noising and fuzzy image segmentation. The main focus of the analysis is the brain image, as it requires rigorous image assessments for an effective classification and detection of patterns. The second phase of the dissertation work expounds the wavelet thresholding for various sets of images. An in-depth investigation of fuzzy rule based optimizer for adapting the wavelet threshold for effective noise suppression has been examined. In this technique, the threshold is further optimized, based on number of criterion including; the intensity, location and size of the noise burst over the malignant patterns. Therefore the present technique improves the post processing diagnostic of images containing small pattern(s) hidden under noise bursts, which otherwise goes undetected. The third phase of the dissertation work studies the impact of non-homogeneous noise on the performance of fuzzy image clustering algorithm. Various results were analyzed for clean, noisy and de-noised images. The purpose here is to segment the malignant areas of noisy brain MRI for effective tumor detection. Fuzzy rule based optimizer plays an important role for adapting the wavelet threshold for the region of interest. The fuzzy information of image contours and noise burst transformed into crisp control decision signals for adapting the threshold. In addition, it was found that the noisy image with no tumor has a false possibility of detecting benign pattern as malignant area. Other research outcome includes the detection of patterns in an image with invisible noise bursts using Multi-resolution Analysis. The result of this course of action is obtained in the diagonal detail components of multi-level decomposition. The difficulties observed in the prevailing methodology include the limited set of research studies conducted to address the issue of non-homogeneous noise in MR Images and the limited accessibility of real images. A good source of validation is the comparison of the de- noised image with that of clean image. Impact of non-homogeneous noise has been explored using directional wavelet. This analysis demonstrates how adversely, different noise patterns affect the computational performance of curvelets and ridglet. The main outcomes of this technique include the impact of non- homogeneous noise on wavelet and curvelet based de-noising methods. An important attribute of this research, is improved methodology for malignant patterns detection in noisy MR Images. This, in turn, makes possible the better development of image diagnostic tools.