the segmentation of brain tumor and its classification in the early stage is very important for the purpose of diagnosis and treatment. This work introduces a new deep neural network model Lu-Net with less layers, less complexity and very efficient for identifying tumors. The work involves classifying brain magnetic resonance (MR) images from a dataset of 253 images of high pixels into two categories of tumors and non-tumors. MR images are initially resized, cropped, preprocessed, and augmented for accurate and rapid training of deep convolutional neural network (CNN) models. The performance of the Lu-Net model has been evaluated using five types of statistical evaluation matrix accuracy, recall, specificity, F-score and accuracy, and its performance also compared with other two types of model Le-Net and VGG-16. CNN models were trained and evaluated on augmented dataset and tested on untrained datasets. The overall accuracy of Le-Net, VGG-16 and the proposed model is 88%, 90% and 98%, respectively, indicating the superiority of the proposed model.