Optimized Weighted Voting System for Brain Tumor Classification Using MRI Images

1. Problem Statement

Accurate classification of brain tumors (specifically glioma, meningioma, and pituitary tumors) from Magnetic Resonance Imaging (MRI) scans is critical for diagnosis and treatment planning. However, this task faces significant challenges:

• Morphological Variability: Tumors exhibit diverse shapes, textures, and anatomical locations.
• Acquisition Conditions: Variations in MRI acquisition settings affect image quality.
• Model Limitations: Traditional machine learning models (e.g., SVM, KNN) often struggle with complex feature extraction, while standalone deep learning models may lack robustness or fail to generalize across different datasets.
• Data Imbalance: Real-world datasets often suffer from class imbalance, complicating model training.

The paper aims to overcome these limitations by proposing a hybrid ensemble learning framework that leverages the complementary strengths of both traditional machine learning and deep learning architectures.

2. Methodology

The proposed system follows a multi-stage pipeline involving data preprocessing, diverse model training, and an optimized weighted voting mechanism.

A. Data Preprocessing and Representation

The system utilizes two public datasets: Figshare (3,064 images, imbalanced) and Kaggle (7,023 images, balanced). To enhance feature extraction, three distinct data representations are generated:

1. Grayscale Images: Original 512x512 MRI scans processed with Balance Contrast Enhancement (BCET) to improve tumor visibility.
2. Edge Images: A pipeline involving BCET, K-means clustering (for skull/soft tissue segmentation), and Canny edge detection (using Gaussian filtering and Sobel operators) to refine tumor boundaries.
3. Handcrafted Features: Histogram of Oriented Gradients (HOG) is extracted to capture global shape and texture structures, chosen for its resilience to contrast changes compared to other descriptors like SIFT or LBP.

B. Classifier Ensemble

The ensemble integrates seven distinct classifiers, each trained on specific input types to capture different feature perspectives:

• Traditional ML:
  • KNN: Trained on HOG features using Euclidean distance.
  • SVM: Trained on HOG features using various kernels (Linear, RBF, etc.).
• Deep Learning (CNNs):
  • CNN-MRI: A custom 3-layer CNN trained on full-resolution grayscale images.
  • ResNet50: Trained on edge-detected images to focus on structural boundaries.
  • Xception: Trained on edge-enhanced images using depthwise separable convolutions.
  • DenseNet121: Trained on grayscale images to leverage dense connectivity for feature reuse.
  • ResNet101: A deeper architecture trained on grayscale images for hierarchical feature extraction.

C. Optimized Weighted Voting System

Instead of simple majority voting, the system employs a weighted voting mechanism where the final prediction ($y_{final}$) is determined by:
$$y_{final} = \arg \max_{c} \left( \sum_{i=1}^{n} w_i \cdot p_{i,c} \right)$$
Where $w_i$ is the weight assigned to classifier $i$ based on its individual accuracy, and $p_{i,c}$ is the confidence score for class $c$.

• Optimization: The weights (ranging from 0 to 10) were optimized using automated tools (AutoKeras for CNNs, Optuna for others) to find the configuration that maximizes ensemble accuracy while excluding underperforming models (weight = 0).

3. Key Contributions

• Hybrid Integration: Successfully combines traditional feature-based methods (HOG + SVM/KNN) with state-of-the-art deep learning architectures (ResNet, DenseNet, Xception) into a single cohesive framework.
• Multi-Representation Learning: Demonstrates that training models on different image representations (raw, edge-enhanced, and feature-extracted) significantly boosts ensemble performance compared to using a single input type.
• Optimized Weighting Strategy: Moves beyond static or equal voting by dynamically assigning weights based on individual model performance, ensuring high-accuracy models (like ResNet101 and DenseNet121) have a dominant influence.
• Robustness Across Datasets: Validates the approach on both balanced (Kaggle) and imbalanced (Figshare) datasets, proving generalizability.

4. Experimental Results

The system was evaluated on the Figshare and Kaggle datasets, achieving state-of-the-art results:

• Comparison: The proposed method outperformed existing single-model approaches (e.g., 23-layer CNN, EfficientNet) and previous ensemble systems (e.g., AlexNet/VGG ensembles) which typically peaked around 97–98%.
• Class Performance: The system achieved near-perfect classification for "No Tumor" and "Pituitary" classes on the Kaggle dataset, with high precision and recall across all categories.

5. Significance and Conclusion

This research highlights the potential of ensemble learning in medical image analysis. By strategically combining diverse algorithms and optimizing their voting weights, the system achieves a level of accuracy (>99%) that is critical for clinical decision-making.

• Clinical Impact: The high accuracy and robustness suggest the framework is scalable and reliable for real-world brain tumor diagnosis, potentially reducing misdiagnosis rates.
• Future Directions: The authors note that while performance is high, distinguishing between glioma and meningioma remains slightly challenging. Future work will focus on dynamic weighting strategies and advanced feature engineering to further refine these specific distinctions.

In summary, the paper presents a highly effective, scalable solution for brain tumor classification that leverages the synergy between traditional and deep learning techniques, validated by superior performance on standard medical imaging benchmarks.