Theses and Dissertations
Permanent URI for this collectionhttp://ir.daiict.ac.in/handle/123456789/1
Browse
2 results
Search Results
Item Open Access Techniques for denoising brain magnetic resonance images(Dhirubhai Ambani Institute of Information and Communication Technology, 2016) Phophalia, Ashish; Mitra, Suman K.Advances in the computational science joined medical imaging domain to help humanity. It offers great support in clinical practices where automatic Computer Added Systems (CAD) help in identification and localization of abnormal tissues. In recent decades, a lot of research tuned non-invasive techniques have been devised to serve mankind. One of them is Magnetic Resonance Imaging (MRI) which provides structural information at higher resolution even in presence of bone structures in the body. Although it is free from ionizing ingredient, factors like electronic circuitry, patient movement etc. provoke some artifacts in imaging system considered as noise. One needs to get rid of these artifacts by means of software processing to enhance the performance of diagnostic process. This thesis is also an attempt to deal with noisy part of MRI and comply with preserving image structures such as boundary details and preventing over-smoothing. It has been observed that, in case of MR data, noise follows Rician distribution. As opposed to additive Gaussian noise, Rician noise is signal dependent in nature due to MR image acquisition process. The thesis constitutes a relationship between MRI denoising and uncertainty model defined by Rough Set Theory (RST). RST already has shown some promising outcomes in image processing problems including segmentation, clustering whereas not much attention has been paid in image restoration task. The first part of the thesis proposes a novel method for object based segmentation and edge derivation given the noisy MR image. The edges are closed and continuous in nature and segmentation accuracy turns out to be better than well-known methods. The prior information is used as cues in various image denoising frameworks. In Bilateral filter framework along with spatial and intensity cues, a new weighing factor is derived using prior segmentation and edge information. This further extends to non local framework where waiver in spatial relation conceded to access similar information from far of neighbors. Under non locality paradigm, a clustering based method is proposed which clubs together similar patches based on similarity criteria. The proposed clustering method uniquely defines clusters of patches under multiple class set up. These clusters are then used to define the basis vectors using Principal Component Analysis (PCA) and Kernel Principal Component Analysis (KPCA) method followed by hard thresholding shrinkage procedure. Afterwards, multiple estimations of a pixel are averaged by number of estimations. In total, number of PCA or KPCA operations are far less than other contemporary methods which repeat the same process over chunks of patches in the image space. The concept is then extended for 3D MRI data. The 3D imaging provides better view of objects from three directions as compared to 2D imaging where only one face of object can be viewed. It involves a complex relationship as compared to 2D imaging and hence is computationally expensive. But it also includes more information which helps in visualizing the object, its shape, boundary etc. similar to real world phenomenon. We extended the segmentation and edge derivation mechanism to 3D data in last part of the thesis. Clustering process is also extended by converting each voxel to one dimensional vector. This part explores various kernels over Rician noise distributed MR data. The results are promising in terms of structure measures even with some simple kernels.Item Open Access Classification of 3D volume data using finite mixture models(Dhirubhai Ambani Institute of Information and Communication Technology, 2010) Phophalia, Ashish; Mitra, Suman K.The 3D imaging provides better view of objects from three directions as compared to 2D imaging where front face of object can be viewed only. It involves a complex relationship as compared to 2D imaging and hence computationally expensive also. But it also includes more information which helps in visualizing the object, its shape, boundary similar to real world phenomenon. The segmentation method should take care of 3D relationship that exists between voxels. The multi channel 3D imaging provides exibility in changing voxel size by changing echo pulse signals which helps in analysis of soft tissues. The application of 3D imaging in MRI brain images help in understanding more clearly the brain anatomy and function. Mixture model based image segmentation methods provide platform to many real life segmentation problems. Finite Mixture Model (FMM) segmentation techniques have been applied in 2D imaging successfully. But these methods do not involve spatial relationship among neighboring pixels. To overcome this drawback, Spatially Variant Finite Mixture Model (SVFMM) was given for classification purpose. In the medical imaging, the probability of noise is high due to environment, technician expertise level, etc. So, a robust method is required which can reduce the noise effect of the images. The Gaussian Distribution is more preferred in the literature but it is not robust against the noisy data. The Student's t Distribution uses Mahalanobis squared distance to reduce the effect of outlier data. A comparative study has been presented between these two distribution functions. In Medical Imaging, segmentation procedures provide facility to separate out different type of tissues instead of manual processing which requires time and efforts. The segmentation methods automate this classification procedure. To reduce the computation time in 3D medical imaging, a sampling based approach called Column Sampling is used. The variance of a column is taken as a measure in sample selection. A comparison is presented for time taken in sample selection from whole volume with Random Sampling. The selected samples are provided to the estimation technique. The parameters of mixture model are estimated using Maximum Likelihood Estimation and Bayesian Learning Estimation in the presented work. The method for estimating parameters of SVFMM using Bayesian Learning is proposed. The Misclassification Rate (MCR) is used for quantitative measure among these methods. This work analyzes FMM and SVFMM models with different probability distribution over two different estimation techniques. The MCR and computational time are considered as quantitative measures for performance evaluation. The different sampling percentage is tried out to estimate the parameters and their MCR and computational time are presented. In conclusion, Bayesian learning estimation SVFMM using Student's t distribution gives comparatively better results.