Performance Comparison of Ensemble Learning Models for Brain Tumor Detection on Augmented MRI Datasets

Abstract

Brain tumors are highly fatal diseases, making early detection a critical factor in improving patient survival rates. Magnetic Resonance Imaging (MRI) has become a primary tool in brain tumor diagnosis; however, manual analysis processes are often time-consuming and prone to subjective errors. This study employs a machine learning-based classification model to detect four categories of brain tumors—glioma, meningioma, pituitary, and healthy—with high accuracy. The methods include image segmentation using the U-Net model, which excels in medical image analysis due to its encoder-decoder architecture with skip connections, allowing efficient integration of spatial and contextual information. Features are extracted using HuMoments, known for their invariance to rotation, translation, and scale, ensuring robust spatial pattern representation. Data normalization is conducted using Robust Scaling and L2 Normalization to address outliers and harmonize feature scales, enhancing model performance. The MRI dataset, originally comprising 7,023 images, was augmented to 8,000 images using techniques such as rotation, flipping, and contrast adjustments to improve class balance and minimize overfitting. Three ensemble algorithms—Random Forest, XGBoost, and Stacking—were employed to train the models, with performance evaluation based on accuracy, ROC-AUC, F1-score, and confusion matrix. The results demonstrate that Random Forest achieved the best performance with an accuracy of 72% and an ROC-AUC of 0.91. This study illustrates the potential of machine learning approaches for automated brain tumor diagnosis, with further improvement possible through model optimization and the use of more diverse datasets.