Abstarct: The automatic segmentation of brain images is a fundamental part in many medical treatments and investigations including diagnosis, locating and tracing of tumor growth, delineation of unhealthy tissues and deciding over the type of medical treatment plan. The extraction of tumors is commonly done manually by radiology experts and is known to be very tieresome and erroneous. Hence the automation of segmentation has become the subject of interest in many medical studies. Automatic segmentation of brain tissues involves many challenges due to the diversity of the location and general shape of brain tumors, their diverse range of intensities or the ambiguity of tissue bounds and partial voluming along the boundaries. Some deficiencies in RF coil also can distort the real image values. Robustness of the method against intensity artifacts and noises is inevitable for the sake of its future success. In this thesis we propose an automatic semi-supervised segmentation algorithm using the random forest classifier. Semi-supervised trained classifiers are specifically beneficial in medical applications where providing the classifier with the experts` labeled training data is laborious. The most important feature in any segmentation algorithm is the intensity levels of voxels which can be easily distorted by intensity inhomogeneity and noise. Regarding the particular application of the system, interference of prior knowledge such as spatial constraints or deformable atlases will enhance its accuracy. In order to extract reliable features, we propose to modify the neighborhood term and define a flexible and adaptive neighborhood named the Gossiping-baxsed neighborhood system. The idea is to impose both spatial and intensity constraints to partition the image into primary regions where pixels inside each region are close neighbors and belong to the same class with high probability. Extracting spatial features inside the region bounds prevents irrelevant pixels to mutilate statistical features of the region. The extracted features are to be used with random forest classifier for final segmentation. The RF classifier has rarely been notified in MRI segmentation however the classifier`s traits such as high speed and accuracy, the ability to handle large feature vectors and missing values and the probability- baxsed decision making make it very appropriate to use in the problem of segmentation and modeling ambiguity around the borders. The brain scans segmentation with random forest showed promising results using the gossip-baxsed region growing algorithm compared to the non-adaptive traditional feature extraction. Although being trained in a semi-supervised mode, the final classifier`s segmentation accuracy on about 800 different MRI slides is as high as most of recently proposed brain segmentation methods which means a noticeable increase in accuracy regarding the volume of testing dataset.