Brain Tumor Detection and Classification Using Deep Learning

Overview

This project presents a deep learning–based system for automated brain tumor detection and classification from 2D brain MRI images. The model determines whether a tumor is present and, if so, identifies the tumor type. The work emphasizes methodological rigor, proper dataset handling, and reliable evaluation suitable for research and academic review.

---

Classification Targets

The model performs multi-class classification with the following categories:

• Glioma
• Meningioma
• Pituitary tumor
• No tumor

---

Dataset

This project uses a publicly available 2D brain MRI dataset sourced from Kaggle.
All images are pre-extracted 2D slices, eliminating the need for volumetric processing or slice generation.

Dataset Characteristics

• Imaging modality: Brain MRI (2D)
• Image format: JPEG/PNG
• Balanced class distribution across four categories
• Labeled and pre-processed for supervised learning

The dataset is commonly used in academic benchmarking and provides a reliable foundation for evaluating multi-class brain tumor classification models.

---

Data Handling and Preprocessing

• Images are resized and normalized to ensure consistency across samples
• Data is split into training, validation, and test sets
• Class balance is preserved across splits
• No data augmentation techniques that could introduce label leakage are applied during testing

---

Model Architecture

Two modeling strategies are implemented:

1. Convolutional Neural Network Trained From Scratch

A custom CNN designed specifically for brain MRI classification.
This approach learns features directly from the dataset without pretrained priors.

2. Transfer Learning with ResNet-18

A pretrained ResNet-18 architecture fine-tuned on the MRI dataset.
Transfer learning significantly improves convergence speed and final classification performance.

---

Training Configuration

• Framework: PyTorch
• Device: CPU
• Loss Function: Cross-Entropy Loss
• Optimizer: Adaptive gradient-based optimizer
• Early stopping based on validation accuracy

---

Evaluation Metrics

Performance is evaluated using:

• Accuracy
• Precision
• Recall
• F1-score
• Confusion matrix analysis

All metrics are reported on a strictly held-out test set.

---

Results

From-Scratch Model

• Test accuracy consistently between 94% and 98%
• Strong class-wise performance across all tumor types

Pretrained ResNet-18 Model

• Final test accuracy: 98.98%
• Near-perfect precision and recall across all classes
• Minimal inter-class confusion observed

---

Reproducibility

• Deterministic training with fixed random seeds
• Clear separation of training, validation, and test data
• Model weights excluded from version control
• Fully reproducible training and evaluation scripts

---

Ethical and Research Considerations

This project is intended solely for research and educational purposes.
It is not a clinical diagnostic system and must not be used in medical decision-making.

---

Future Work

• Evaluation on multi-institutional datasets
• Extension to 3D volumetric MRI classification
• Tumor segmentation and localization
• Model explainability and interpretability analysis

---

Disclaimer

All results are experimental and reflect performance on a specific public dataset.
Clinical applicability requires extensive validation and regulatory approval.