Medical images often require segmenting into different regions in the first analysis stage. Relevant features are selected to differentiate various regions from each other, and the images are segmented into meaningful (anatomically significant) regions based on these features. The purpose of this study is to present a model for segmenting and identifying the local tumor formation in MR images of the human brain. The proposed system operates in an unsupervised manner to minimize the intervention of expert users and to achieve an acceptable speed in the tumor classification process. The proposed method includes several steps of preprocessing for different brain image classify that Perform the normalization task. These preprocessing steps lead to more accurate results in high-resolution images and ultimately improve the accuracy and sensitivity of tumor separation from brain tissue. The output of this stage is applied to a self-encoding neural network for image zoning. By nature of self-encoding networks, leads to reduce the dimensionality of tumor pixels from the surrounding healthy environment, which significantly helps remove regions incorrectly extracted as tumors. Finally, by extracting features from the previous stage's output through Otsu thresholding, the surrounding area and type of tumor are also extracted. The proposed method was trained and tested using the BRATS2020 database and evaluated by various performance metrics. The results based on the Dice Similarity Coefficient (DSC) show an accuracy of 97% for the entire MR image and improved detection accuracy compared to other methods, as well as a reduction in the cost of the diagnostic process.