Brain tumours are a critical health concern, with magnetic resonance imaging (MRI) being pivotal in their detection. With the advent of deep learning (DL), convolutional neural networks (CNNs) have showcased exemplary performance in image-based tasks, including medical diagnoses. In our investigation, we classify MRI brain images into four categories: no tumour, glioma, pituitary, and meningioma using CNNs. Recognizing the importance of feature optimization in enhancing model performance, we introduce particle swarm optimization (PSO)—a bio-inspired algorithm known for efficient search and optimization in high-dimensional spaces. Our method integrates CNN-based feature extraction with PSO for optimization, followed by machine learning classification. Preliminary results reveal the potency of this integrated approach. The CNN-PSO-ML framework outstripped benchmark models, achieving an accuracy rate of 98.82%. Among various architectures, VGG16 combined with logistic regression led with an accuracy of 94.56%, followed by InceptionV3 and Resnet50.