Ensemble Learning for MRI-Based Brain Tumor Classification: A Weighted Voting Approach

MRI is essential for detecting and diagnosing brain tumors, where accurately distinguishing glioma, meningioma, and pituitary tumors is vital for effective treatment planning. However, tumors' complex morphology and MRI imaging variations present significant challenges for reliable classification. Deep learning models, particularly Convolutional Neural Networks (CNN) and ResNet architectures, have demonstrated impressive performance in medical image analysis but often struggle with generalization across different datasets. On the other hand, traditional classifiers such as Support Vector Machines (SVM) and K-Nearest Neighbors (KNN) leverage handcrafted features like Histogram of Oriented Gradients (HOG), which can effectively capture structural details but may lack the adaptability required for diverse imaging conditions. This thesis presents an ensemble-based voting system that integrates multiple single classifiers, leveraging their complementary strengths to improve brain tumor classification accuracy. The system combines traditional machine learning and deep learning classifiers with the custom CNN model (CNN-MRI), ResNet101, and DenseNet121 trained on grayscale MRI images. At the same time, ResNet50 and Xception utilize edge-detected images to enhance feature extraction. SVM and KNN using HOG features are also incorporated to capture shape and texture details. A weighted voting mechanism determines the final classification, assigning higher influence to models with greater individual accuracy. Image processing techniques are applied to improve input quality and optimize model learning. Balance Contrast Enhancement (BCET) enhances the color contrast between different MRI components, such as tumors, tissues, skull, and fluid, making key structures more distinguishable. K-means clustering then segments these components, ensuring more precise separation. Finally, Canny edge detection filters out irrelevant pixels, preserving only the edges of essential structures to highlight tumor boundaries more effectively. Experimental evaluations conducted on both the Figshare MRI dataset and the Kaggle MRI public dataset validate the effectiveness of this multi-model ensemble approach. The proposed method has demonstrated exceptional performance, achieving over 99% accuracy across two benchmark datasets. Compared to previous studies using the same datasets, this approach ranks among the most accurate and reliable solutions for brain tumor classification. These results highlight the effectiveness of the ensemble voting system, reinforcing its potential for real-world medical applications where AI-driven diagnostics can enhance accuracy, consistency, and clinical decision-making.