🧠 Brain Tumor Classification using ResNet18

A deep learning project for classifying brain MRI images into four classes using transfer learning with ResNet18 and PyTorch.

📌 Classes

• Glioma
• Meningioma
• No Tumor
• Pituitary

🛠️ Tech Stack

• Python
• PyTorch
• Torchvision
• NumPy
• Matplotlib
• scikit-learn
• Jupyter Notebook

🔬 Approach

The project follows a complete image-classification workflow:

1. Dataset exploration
2. Stratified train/validation split
3. Image preprocessing and augmentation
4. Transfer learning using pretrained ResNet18
5. Training with a frozen backbone
6. Partial fine-tuning of ResNet18
7. Model evaluation using accuracy, precision, recall and confusion matrices
8. Misclassification analysis

Three ResNet18 configurations were evaluated:

Model	Test Accuracy
Frozen ResNet18	80.69%
Layer4 + FC Fine-tuning	91.69%
Layer3 + Layer4 + FC Fine-tuning	93.19%

The final model achieved 93.19% accuracy on the 1,600-image test set.

📂 Project Structure

Directory structure:
└── Brain-Tumor-Classification/
    ├── images/
    │   ├── CompleteFT CM.png
    │   ├── CompleteFT Metrics.png
    │   ├── Frozen Backbone CM.png
    │   ├── Frozen Backbone Metrics.png
    │   ├── Misclassified Images.png
    │   ├── PartialFT CM.png
    │   └── PartialFT Metrics.png
    ├── models/
    │   ├── ResNET_Frozen.pt
    │   ├── ResNET_Frozen_FT.pt
    │   └── ResNET_Frozen_PartialFT.pt
    ├── notebooks/
    │   ├── 00. Data Exploration.ipynb
    │   ├── 01. ResNET.ipynb
    │   └── 02. ResNET Evaluation.ipynb
    ├── README.md
    ├── requirements.txt
    ├── split.py
    └── src/
        ├── data.py
        ├── engine.py
        ├── evaluation.py
        ├── metrics.py
        └── plots.py

🚀 Getting Started

1️⃣ Clone the repository

git clone <YOUR_REPOSITORY_URL>
cd Brain-Tumor-Classification

2️⃣ Create a virtual environment

python -m venv .venv

Activate it:

Windows:

.venv\Scripts\activate

Linux/macOS:

source .venv/bin/activate

3️⃣ Install dependencies

pip install -r requirements.txt

4️⃣ Prepare the dataset

Place the original dataset inside the archive/ directory:

archive/
├── Training/
│   ├── glioma/
│   ├── meningioma/
│   ├── notumor/
│   └── pituitary/
└── Testing/
    ├── glioma/
    ├── meningioma/
    ├── notumor/
    └── pituitary/

Run the dataset splitting script:

python split.py

This creates the following structure:

dataset/
├── train/
├── val/
└── test/

The original test set is kept untouched.

5️⃣ Run the notebooks

Launch Jupyter:

jupyter notebook

Open the notebooks inside the notebooks/ directory to reproduce the analysis and experiments.

📦 Pretrained Models

Trained ResNet18 checkpoints are available in the models/ directory.

The final model is the Layer3 + Layer4 + FC fine-tuned ResNet18.

⚠️ Disclaimer

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

---