Brain Tumor Classification

CNN-based classification of brain tumor types from MRI scans.

Dataset

Source: Brain Tumor MRI Dataset (Kaggle)

4 classes: glioma, meningioma, notumor, pituitary — each split into Training/ and Testing/.

Model Architecture

SimpleCNN with 10.7M parameters:

• 3 Convolutional blocks (32, 64, 128 filters)
• Batch Normalization + Dropout
• 2 Dense layers (256, 128 neurons)
• Softmax output (4 classes)

Results

Configuration	Accuracy	Loss	Change
Baseline	42.39%	3.2599	-
Fine-tuned	46.70%	2.8472	+4.31%

Fine-tuned hyperparameters: LR=0.0005, Batch=16, Epochs=40

Per-Class Performance (Fine-tuned)

Class	Precision	Recall	F1-Score
Glioma	1.00	0.14	0.24
Meningioma	1.00	0.21	0.35
No Tumor	0.37	0.92	0.53
Pituitary	0.54	0.66	0.59

Installation

Step 1: Install UV

curl -LsSf https://astral.sh/uv/install.sh | sh

Step 2: Setup Project

uv sync

Step 3: Download Dataset

chmod +x download_dataset.sh
./download_dataset.sh

The script checks UV, installs Kaggle CLI, prompts for credentials, and downloads to ./data/.

Manual alternative:

uv add kaggle
mkdir -p ~/.kaggle
# Create ~/.kaggle/kaggle.json: {"username":"your_username","key":"your_api_key"}
chmod 600 ~/.kaggle/kaggle.json
uv run kaggle datasets download -d masoudnickparvar/brain-tumor-mri-dataset -p ./data --unzip

Usage

make visualize   # generate data exploration images
make train       # train model (~2 hours, saves to models/)
make publish     # push to GitHub + HuggingFace

Individual commands

uv run python scripts/data_exploration.py
uv run python scripts/train_models.py
uv run python scripts/predict_single_image.py data/Testing/glioma/Te-gl_1.jpg

Model Weights

Weights hosted on HuggingFace: supakornn/Brain-Tumor-Classification

# Download model weights
git clone https://huggingface.co/supakornn/Brain-Tumor-Classification hf-model
cp hf-model/model/*.h5 models/

Limitations

• Low accuracy (46.70%) insufficient for clinical use
• Poor glioma detection (14% recall)
• Simple architecture inadequate for medical imaging

License

MIT License — Educational and research purposes.