Unusual masses of tissue where the indefinite growth of cells, is uncontrolled by the normal cell control mechanisms, are known as intracranial tumors or brain tumors. Brain tumors are prevalent at a rate of 5–10 per 100,000 populations in India. At least 28,000 cases are registered annually by the International Association for the Registry of Cancer (IARC); more than 24,000 deaths occur every year. There is an absolute need for a system to identify and classify tumors automatically to assist radiologists and physicians. Nevertheless, the accuracy of the present techniques must be upgraded for necessary treatments. In the following research, the capability comparison of the CNN model is presented, and the transfer learning approach for models like VGG-16, EfficientNet-B3, and Inception v3, for the analysis carried out over the dataset made up of training images counting to a total of 2902 MRI images and testing images of total 394 images. The mentioned dataset was classified into 3 types: glioma, meningioma pituitary tumors, and no tumor. The experimental results show that EfficientNet-B3 was the most efficient model and it was 98.5% accurate in training and 97% accurate in testing. A decision-making tool for studying brain tumor diagnostic tests can be developed using the proposed EfficientNet CNN architecture because of its high accuracy and favorable F1-score.