The human brain tumour is an abnormal growth of brain cells which is a fast-spreading deadly disease in the world and targets adults and children. The symptoms and survival rates entirely depend on type of the tumour, location, grade and size. The rapid tumour classification is essential to increase the survival rate of the patients. Magnetic resonance imaging (MRI) gives more soft tissue characteristics about tumours but manual classification from MR images is a time-consuming process. Rapid growth of deep learning models will give automatic methods for tumour type classification. This manuscript proposes a hybrid method named MResNet50-SVM. It is used for brain tumour classification into three labels: glioma, meningioma and pituitary. The ResNet50 architecture is altered with 183 layers and 197 connections to collect the image features. Extracted features were trained with a support vector machine (SVM) for tumour classification. The major advantage of the proposed method is it hybrid the neural network and machine learning concept to improves the accuracy. The method works with two online repositories Kaggle, Figshare and real-time patients’ datasets were collected from Meenakshi Mission Hospital and Research Centre (MMHRC), Madurai. The performance of the proposed work is computed using sensitivity, specificity, positive predicted value (PPV), and negative predicted value (NPV), accuracy and corresponding obtained values are 99.26, 99.62, 99.25, 99.62, and 99.5%. The classification accuracy of current work is compared with pre-trained DL models like GoogLeNet, ShuffleNet, SqueezeNet and ResNet18. It also compared with four state-of-the-art methods. Finally, the proposed MResNet50-SVM proved a promising classification accuracy than pre-trained models and state-of-the-art-methods.