Intracranial tumors are characterized by abnormal tissue growth within the human brain. As the brain is a critical organ, uncontrolled growth in it can cause serious damage. Earlier its crucial to diagnose and predict tumor conditions, mitigate complications and improve treatment outcomes. Recent advancements in this domain have extensively leveraged deep learning. We utilize deep learning and machine learning models to classify magnetic resources. We can classify magnetic resonance imaging (MRI) scans into four categories: meningiomas, gliomas, pituitary adenomas, and non-tumorous conditions. However since then, many existing studies have relied on complex architectures to achieve high classification accuracy. This research proposes a novel, customized convolutional neural network (CNN) architecture for the efficient and accurate classification of intracranial tumors. The designed CNN consists of nine layers that balance simplicity and performance. By reducing the number of layers, it achieves a higher performance compared to traditional machine learning models. The proposed approach achieves computational efficiency in the models without compromising accuracy. The experimental results show that the model achieves an outstanding classification and the accuracy was 99.41%, with a minimal loss of 0.0198%. The proposed CNN model achieved 99.41% accuracy on the single-source dataset. However, when evaluated on a multi-source external dataset, accuracy dropped to 73.84%, revealing generalization challenges. This suggests the need for domain adaptation techniques to improve robustness across different MRI scanners.