The main objective of this research work is to develop, test and evaluate an identification support system that is able to provide accurate, fast and reliable diagnosis of brain tumor in MRimages. Keeping in consideration that human decision making skills are mainly dependent on experience and prone to error due to fatigue, Artificial Intelligence (AI) can be utilized as an effective aid in the field of medicinal sciences for tumor diagnosis through image recognition. Therefore, this thesis strives to develop such an intelligent system that can be used for the segmentation and classification of infiltrative brain tumors known as Low Grade and High Grade in MR images. In order to tackle the complex task of brain tumor segmentation in MR images, we present an adaptive algorithm that formulates an energy based stochastic segmentation with a level set methodology. This hybrid technique efficiently matches, segments and determines the anatomic structures within an image by using global and local energies. After evaluating the algorithm on low and high grade images, it was noted that there was an improvement in the resultant similarity between segmented and truth (original) images. Once effective segmentation was achieved we could then work on the next step of tumor identification; classification. In the second part of the process we proposed two classification frameworks, machine learning and deep learning. In machine learning, we first extracted 22 probabilistic features using gray level co-occurrence matrix methodology that served as input features for the classifiers. Then we showed the improvement in classification (through machine learning) accuracy by providing two methodologies in which the first one involved v classification directly after feature extraction whereas in the second we reduced the extracted features using principal component analysis and then applied those reduced features to several classifiers. The second framework that we proposed was the brain tumor classification of segmented MR images through optimized CNN-Deep belief learning model. It scales to various image sizes by distributing the hyper-parameters and weights among all locations in an image. The presented model is translation invariant and is compatible with top-down and bottom-up probabilistic inference. This hierarchical classifier was optimized by regularization, that mitigates the effect of overfitting for small datasets, stochastic gradient decent, which works efficiently by utilizing only a small set of samples from a whole training set to infer the gradient and fine tuning of constraints. A comparative analysis, based on accuracy, error/loss and computation time, was carried out between the pre-processed non-segmented and segmented MR images after classification was completed. The results showed that the accuracy of proposed optimized CNN-deep belief learning classifier with segmented MR images was higher while the loss and execution time were reduced. These methodologies transcend the confines of MR image processing due to their effective modularity allowing them to be suitable for other medical imaging and computer vision tasks.