Hybrid Tumor Classification Using Vision Transformers and PSO-Optimized Feature Selection with XGBoost

Abstract

To enhance patient outcomes and efficiently direct treatment, it is critical to quickly and precisely classify brain lesions in magnetic resonance imaging (MRI). Our hybrid classification system combines an XGBoost classifier, Principal Component Analysis (PCA)-driven dimensionality reduction, Particle Swarm Optimization (PSO)-based feature selection, and Vision Transformer (ViT)-based feature extraction in a way that complements each other. We evaluate our approach using the Brain Tumor Classification (MRI) dataset from Kaggle, which consists of 3,264 T1-weighted images evenly distributed across four categories: healthy, meningioma, pituitary tumor, and glioma. To preserve 95% of the variance, the ViT encoder compresses the large amount of global context data it receives using PCA. By selecting the most valuable components, PSO further enhances this representation and reduces the feature count by over 70%. The final classification accuracy of 97.8%, 98% precision and recall, and 0.99 area under the ROC curve (AUC) are achieved by XGBoost using the optimized feature subset. With a mean accuracy of 96.5% (±1.0%) and a mean AUC of 0.985 (±0.005), five-fold cross-validation demonstrates the stability of the model. In addition to being more effective and simpler to comprehend, our pipeline performs on par with or better than the top CNN and transformer models. This study demonstrates that there is a powerful, scalable method for automatically diagnosing brain tumors using sophisticated deep learning architectures and evolutionary optimization. Other medical imaging applications might also make use of this technique.