The classification and diagnosis of brain tumours remain reliant on the histological analysis of biopsy specimens. The present approach is intrusive, laborious, and prone to human error. These limitations emphasize the necessity of adopting a fully automated approach for multi-classification of brain tumours using deep learning. The objective of this article is to detect feature sets which are further optimized to get better feature maps to convolutional neural network (CNN). 25 layers consisting of six stage convolution layers are used to extract the feature sets of brain tumour images in the proposed CNN model. This network architecture enables less computational complexity compared to conventional CNN models in the classification of brain tumours into four classes, namely pituitary, meningioma, glioma, and no tumour. The proposed CNN model is assessed using quality metrics, namely sensitivity, accuracy, specificity, precision, and F1 score against leading-edge models including GoogLeNet, VGG-16, ResNet-50, Inceptionv3, and AlexNet. Large and publicly accessible clinical datasets yield satisfactory categorization outcomes. The proposed CNN model can assist radiologists and physicians in validating the initial screening for the purpose of classifying brain tumours.