A Deep Learning-Based Brain Tumor Classification System Using MRI Images from Bangladesh 

Technical Summary: A Deep Learning-Based Brain Tumor Classification System Using MRI Images from Bangladesh

Problem Statement
Brain tumors are critical neurological disorders requiring accurate and early diagnosis for effective treatment. While Magnetic Resonance Imaging (MRI) is the standard for detection, manual interpretation is time-consuming, subjective, and dependent on radiologist expertise. Existing deep learning (DL) solutions often face two significant limitations: a lack of region-specific training data, particularly for developing nations like Bangladesh where imaging protocols and demographics differ from global datasets, and the "black box" nature of DL models which hinders clinical trust due to a lack of interpretability.

Methodology
The study proposes a deep learning framework utilizing transfer learning with a DenseNet-121 architecture to classify brain MRI scans into four distinct categories: Glioma, Meningioma, Pituitary, and Normal.

• Dataset: The system was trained on a Bangladeshi Brain Cancer MRI dataset containing 1,600 raw images (400 per class). These were augmented to 6,000 images. All images were resized to 224 × 224 × 3 to match the input requirements of the pre-trained DenseNet-121 model.
• Architecture: The model employs a transfer learning approach where the top layers of the pre-trained DenseNet-121 are removed. A custom classification head is added, consisting of a GlobalAveragePooling2D layer, a Dropout layer (0.5) for regularization, and a Dense layer with 4 units using SoftMax activation.
• Training: The model was trained using the Adam optimizer and categorical cross-entropy loss function. The dataset was split into training, validation, and test sets (80/10/10).
• Explainability (XAI): To address the interpretability gap, the study integrates four visualization techniques: Grad-CAM, Grad-CAM++, LayerCAM, and ScoreCAM. These generate heatmaps to highlight the specific regions in the MRI scans that influenced the model's predictions, ensuring the model focuses on medically relevant anomalies rather than background noise.
• Deployment: The system was deployed as a functional web application (using FastAPI and Docker) that provides real-time classification with confidence scores and overlays the XAI heatmaps on the input images.

Key Results
The proposed system demonstrated robust performance on the test set:

• Accuracy: The model achieved a test accuracy of 96.52%.
• Class-wise Performance: The classification report indicates high consistency across all classes with a balanced test set (151 images per class):
  • Glioma: Precision 0.95, Recall 0.95, F1-score 0.95.
  • Meningioma: Precision 0.96, Recall 0.95, F1-score 0.96.
  • Normal: Precision 0.96, Recall 0.99, F1-score 0.97.
  • Pituitary: Precision 0.99, Recall 0.97, F1-score 0.98.
• Metrics: The system achieved a Macro Average and Weighted Average of 0.97 for precision, recall, and F1-score.
• Comparative Analysis: The study compares its results against recent literature, including multiscale CNNs with LSTM (89.5%), ResNet-50 variants (92%), and ensemble methods (up to 94.8%). The proposed DenseNet-121 model outperformed these existing methods, achieving the highest accuracy (96.52%).
• Visual Validation: Confusion matrices and ROC/PR curves confirmed low misclassification rates and strong discrimination capabilities. The XAI heatmaps visually confirmed that the model's attention was correctly aligned with tumor regions.

Significance and Claims
The paper claims that its primary contribution is the development of a high-accuracy, interpretable diagnostic tool specifically trained on indigenous Bangladeshi data, addressing the generalization gap found in models trained on datasets from developed countries. By integrating multiple XAI techniques, the study aims to bridge the gap between high-performance AI and clinical trust, providing radiologists with visual evidence to verify the model's decisions.

The authors position this work as a step toward reliable AI-assisted diagnostic tools for regions with limited access to specialized neurological expertise. They emphasize that the combination of high accuracy (96.52%) and transparency via heatmaps makes the system a viable candidate for clinical decision support, though they note that future work is required to expand dataset diversity and integrate multimodal patient data for broader clinical validation.