# 🧠 BRATS 2020 Brain Tumor Segmentation with MONAI
This repository contains a **baseline 3D UNet pipeline** built using [MONAI](https://monai.io/) for the **BRATS 2020 dataset**.
It demonstrates preprocessing, training, and visualization workflows for multi-class brain tumor segmentation.
kaggle Project link: https://www.kaggle.com/code/hassassinsp/monaithon-2k25-brats
---
## 📌 Project Overview
- **Task**: Segment brain tumors from 3D MRI scans.
- **Dataset**: [BRATS 2020](https://www.kaggle.com/datasets/awsaf49/brats20-dataset-training-validation) (369 training cases).
- **Frameworks**: MONAI + PyTorch.
- **Model**: Baseline **3D UNet** with Dice Loss.
---
## ⚙️ Pipeline
1. **Preprocessing**
- Spacing, Orientation, Normalization
- Random cropping, flips, rotations
- Custom augmentations: Gaussian noise
2. **Training**
- 3D UNet (Residual Units)
- Dice Loss + Adam Optimizer
- Mini-batches on GPU
3. **Evaluation** (planned)
- Sliding window inference
- Post-processing (morphological ops, connected components)
- Metrics: Dice, Hausdorff95, Sensitivity, Specificity
4. **Deployment** (planned)
- Streamlit/Gradio interactive app
---
## 🚀 Quick Start (Kaggle)
```bash
!pip install --no-deps monai nibabel -q
```import monai
print("MONAI:", monai.__version__)- Add dataset in Kaggle:
awsaf49/brats20-dataset-training-validation - Run notebook cells to train & visualize.
(All images generated from training cases in BRATS 2020)
We also visualized segmented brain tumors in 3D using 3D Slicer.
This provides an intuitive view of tumor location, size, and spread across the brain volume.
- Advanced architectures: UNet++, SegResNet, Swin UNETR.
- Better augmentations (elastic deformation, histogram matching).
- Full evaluation with ROC/PR curves.
- Deploy demo app with Streamlit/Gradio.
Hackathon Team — Ctrl+Alt+Heal