The identification and classification of tumors in the human mind from MR images at an early stage play a pivotal role in diagnosis such diseases. This work presents the novel Deep Neural network with less number of layers and less complex in designed named U-Net (LU-Net) for the detection of tumors. The work is comprised of classifying the brain MR images into normal and abnormal class from the dataset of 253 images of high pixels. The MR images were 1st resized, cropped, preprocessed, and augmented for the accurate and fast training of deep neural models. The performance of the Lu-Net model is evaluated using five types of statistical assessment metrics Precision, Recall, Specificity, F-score, and Accuracy, and compared with the other two types of model Le-Net and VGG-16. The CNN models were trained and tested on augmented images and validation is performed on 50 untrained data. The overall accuracy of Le-Net, VGG-16 and Proposed model received were 88%, 90%, and 98% respectively.