Efficient Hierarchical Graph-Based Segmentation for Tumor Detection in Magnetic Resonance Images

Abstract

Segmentation of medical images hold a very important position where prior to surgery and post-surgery decisions need to be taken for initializing the recovery process and to reduce the time to be taken in recovery. Tumor extraction and its close analysis are very challenging tasks in medical image processing. Manual segmentation is not ideal for abnormal tissues when compared to high speed computing machines to visually note the location of tumors, their size and volume. Also, detection of tumor which is for the specialists is very time consuming manually for which they feel burdened. The most optimal solution for the extraction of the tumor is image segmentation. Plenty of techniques are introduced to segment the image and extract the desired part/object of the image but they require very excessive user intervention and other limitations. In this thesis we present a novel and robust practical algorithm for tumor image segmentation with minimum user intervention/interaction to assist researchers and clinicians in tumor treatment and follow up. This technique has been devised to feature out tumor from a Magnetic Resonance Images MRI. Algorithm is based on parametric probability density function that is Gaussian Mixture Model (GMM). The Parameters in Gaussian Mixture Model are estimated through training data using the Maximum A Posteriori (MAP) from refined prior model. Each Iteration advances the image segmentation depending on the number of components determined by BIC. Considering efficient application of Gaussian Mixture Model on image with the defined colors for the connected components in close vicinity. Bayesian Information Criterion is used as a cost function for model selection semi supervised seeded image segmentation technique Laplacian coordinates (LC) is applied which is directionally dependent and builds graph keeping close the similar data. Proposed technique is efficient and trivial mathematical solution. It out performs all the prior work with excellent visual results and qualitative metrics, and accuracy is 95-99% on brain tumor dataset. The experimental results are promising to enhance the visibility of multi tumoral structures and treatment.