Segmentation of MRI brain images using statistical approaches

The segmentation of brain MRI images is a challenging and complex task, due to noise and inhomogeneity. The Gaussian Mixture Model (GMM) is a clustering algorithm that is commonly used for brain MRI segmentation. Usually, the Markov Random Field (MRF) model is used to capture neighbourhood information to make GMM more robust against noise, which is time-consuming and can be improved. Noise is one of the obstacles for brain MRI segmentation. The non-Local means (NL-means) algorithm is a state-of-the art neighbourhood-based noisereduction method which is time-consuming and its accuracy can be improved. Intensity inhomogeneity (where pixels belonging to the same tissue have different intensities) is another obstacle for brain MRI segmentation. Filterbased methods are commonly used for inhomogeneity correction which is very simple and fast; but in general, inhomogeneity correction algorithms produce an estimation of inhomogeneity field. Therefore, these algorithms can be improved upon. A neighbourhood-based noise-reduction algorithm which uses the edges of an image is proposed. A sample in the neighbourhood of a pixel does not contribute in the grey level estimation if an edge exists between the sample and the pixel. Also, a filter-based image inhomogeneity-correction algorithm is proposed which uses the maximum filter for inhomogeneity field estimation. Moreover, three improvements of EM for brain MRI segmentation are proposed, which incorporate neighbourhood information in a new manner in the clustering process. In addition, two algorithms for the post-processing of clustering results using user-interaction and the re-evaluation of boundary data in each cluster are presented. The Dice similarity index for the proposed noise-reduction algorithm on 9% noise level was obtained about 0.918 with considerable improvement compared to NL-means. The Dice similarity index for the proposed inhomogeneity-correction algorithm on 40% inhomogeneity level was 0.933 and for real images 0.7627 with considerable improvement compared to recent methods. The Dice similarity index for one of the proposed improvements which yields the best results was: without post-processing equals to 0.8211, with re-evaluation of boundary data equals to 0.848 and with user-interaction equals to 0.8415. The proposed improvements with post-processing yield higher similarity index than several state-of-the-art neighbourhood-based extensions for EM and Fuzzy C-Mean (FCM). One of the proposed improvements yields higher similarity index than other competing methods even without post-processing.

Preview PDF FK 2011 27R.pdf Download (1MB) | Preview

Actions (login required)

View Item