Brain Tumor MRI Segmentation

A deep learning web application that segments brain tumors from MRI scans and generates plain-English radiology summaries using a large language model.

Live Demo | Training Notebook

---

Demo

---

How It Works

1. Upload a FLAIR MRI slice (T1ce is optional but will improve accuracy)
2. A trained U-Net model segments the tumor region in real time
3. The predicted tumor boundary is overlaid in red on the original scan
4. An AI-generated radiology summary is produced using the Gemini API. This is in plain English and should be understood by the average person.

---

Model Architecture

• Architecture: U-Net with ResNet34 encoder and scSE attention decoder
• Library: segmentation-models-pytorch
• Input: 2-channel (FLAIR + T1ce), 256×256 axial MRI slices
• Loss: Combined Dice + Binary Cross Entropy
• Dataset: BraTS 2020 (369 patients, ~50,000 valid slices)
• Augmentation: Albumentations pipeline (HorizontalFlip, RandomRotate90, ElasticTransform, RandomBrightnessContrast)

---

Results

Model : Dice Score

1. Baseline U-Net (FLAIR only) = 0.4219
2. U-Net + T1ce modality = 0.4283
3. Attention U-Net (scSE) = 0.4204
4. Augmented model (in progress) = 0.3747*

*Augmented model is still training and the lower score reflects insufficient epochs, not augmentation failure

---

Tech Stack

• Model: PyTorch, segmentation-models-pytorch, Albumentations
• Data: BraTS 2020, nibabel, numpy
• App: Streamlit
• LLM: Google Gemini API
• Model Hosting: Hugging Face Hub
• Training: Kaggle T4 GPU

---

Run Locally

git clone https://github.com/mehanana/brain-tumor-segmentation.git
cd brain-tumor-segmentation
pip install -r requirements.txt

Create a .env file: GEMINI_API_KEY=your_key_here

Run the app:

streamlit run app.py

---

Disclaimer

This tool is for educational and portfolio purposes only. It is not a medical device and should not be used for clinical diagnosis. Always consult a qualified physician for medical decisions.