Advanced Deep-Transfer-Fusion Models for Effective Multi-class Classification of Brain Tumour Using MRI Data

Brain tumours develop when brain cells grow unhealthily and, if not caught in time, can be fatal. Although MRI has several uses in cancer detection, there is some inter-observer variability because manual interpretation is laborious. To tackle these problems, this research presents an artificial intelligence (AI) classification model called Enhanced TriFusion XGB. It can accurately identify multi-class brain tumours in MRI scans. The model combines complementary deep features that are produced by three trained convolutional neural networks, InceptionV3, ResNet-50 and EfficientNetB7, which encode multi-scale patterns, deep residual representations, and fine-grained structure details, respectively. Feature-level fusion is done and classification is done with the help of an XGBoost meta-learner, which has been chosen due to its regularisation and ability to work with high-dimensional representations. The complex preparation pipeline involves scaling the images, converting them to grayscale, using CLAHE contrast enhancement, normalising the intensity, and augmenting the data. SMOTE was employed to solve the problem of class imbalance. This method guarantees that students enroll in the right number of courses in meningioma, glioma, pituitary, and non-tumour subjects. During model training, categorical cross-entropy loss and the Adam optimiser are used. These models perform better in the experiments than other deep learning models like DenseNet, Vision Transformers, VGG19, and conventional boosting techniques. They have excellent accuracy, recall, and F1-score levels with a success rate of around 96.53%. The findings are encouraging for the hybrid model's use in clinical decision support as they demonstrate that it can correctly detect complex tumour traits with minimal false positives. Future work will focus on ameliorating model interpretability, computational efficacy, and adaptability to diverse datasets.