Brain Tumor Segmentation — BraTS 2023

3D brain tumor segmentation on multimodal MRI data using MONAI and PyTorch, with a full MLOps pipeline including experiment tracking, CI/CD, containerized inference, and an interactive 3D demo.

---

Demo

🔗 Live Demo — Hugging Face Spaces

Upload any BraTS-format MRI volume and visualize predicted tumor subregion masks interactively across axial, coronal, and sagittal planes.

---

Results

Evaluated on 188 held-out BraTS 2023 GLI test cases. Metrics computed using standard image-level Dice and Hausdorff95.

Subregion	Dice ↑	HD95 (mm) ↓
Whole Tumor (WT)	0.930	5.94
Tumor Core (TC)	0.910	3.76
Enhancing Tumor (ET)	0.848	3.23

Leaderboard Comparison

Model	WT Dice	TC Dice	ET Dice	WT HD95	TC HD95	ET HD95
This Model (3D U-Net)	0.930	0.910	0.848	5.94	3.76	3.23
BraTS 2023 Winner (ensemble)†	0.901	0.867	0.851	14.94	14.47	17.70
KLE-based SOTA (2024)†	0.929	—	—	2.93	6.78	10.35

† Challenge entries report lesion-wise metrics which penalize missed lesions and false positives, resulting in lower scores than the standard image-level metrics used here.

Dice Score Distribution

Full evaluation details and qualitative visualizations in notebooks/02_evaluation.ipynb.

---

Overview

This project trains a 3D U-Net on the BraTS 2023 GLI dataset to segment brain tumor subregions from multimodal MRI volumes. The model predicts three clinically meaningful subregions:

• Whole Tumor (WT) — full tumor extent (labels 1 + 2 + 3)
• Tumor Core (TC) — necrotic core and enhancing tissue (labels 1 + 3)
• Enhancing Tumor (ET) — actively growing tumor region (label 3)

Beyond model training, the project implements a production-style MLOps pipeline: experiment tracking with MLflow, automated CI/CD via GitHub Actions, containerized inference via Docker, and a publicly accessible live demo.

---

Architecture

• Model: 3D U-Net with residual connections (MONAI)
• Input: 4-channel multimodal MRI volume (T1, T1ce, T2, FLAIR), skull-stripped and intensity normalized
• Output: 3-class voxel-wise segmentation mask (WT / TC / ET)
• Loss: Dice loss + cross-entropy (combined)
• Inference: Sliding window inference with 128³ ROI over full volume

---

Project Structure

brats-tumor-segmentation/
├── .github/workflows/      # GitHub Actions CI/CD (ci.yml)
├── checkpoints/            # Saved model weights (best_model.pth)
├── configs/                # YAML training configuration
├── data/                   # gitignored — BraTS NIfTI volumes
│   ├── raw/                # original downloaded volumes
│   ├── processed/          # skull-stripped, normalized volumes
│   └── splits/             # train/val/test split definitions
├── docker/                 # Dockerfile + .dockerignore
├── notebooks/
│   ├── figures/            # saved evaluation plots
│   ├── 01_data_exploration.ipynb
│   └── 02_evaluation.ipynb
├── src/
│   ├── inference/          # sliding window inference, FastAPI endpoint
│   ├── interface/          # NiiVue 3D viewer (index.html + sample cases)
│   ├── preprocessing/      # skull stripping, normalization, split creation
│   ├── training/           # model, dataloader, transforms, training loop
│   └── utils/              # data download, manifest utilities
├── tests/                  # pytest unit tests
├── .env
├── pyproject.toml
├── requirements.txt
└── README.md

---

MLOps Pipeline

Component	Tool	Details
Experiment tracking	MLflow	Logs hyperparameters, per-epoch metrics, and segmentation visualizations as artifacts
Model registry	MLflow	Version tags with metric annotations
CI/CD	GitHub Actions	Linting, unit tests, and model checks on every PR
Containerization	Docker	Reproducible inference environment
Deployment	Hugging Face Spaces	Publicly accessible live demo

---

Setup & Reproducibility

Requirements

• Python 3.11
• See requirements.txt for full dependency list

Installation

git clone https://github.com/AndrewVFranco/brats-tumor-segmentation.git
cd brats-tumor-segmentation
python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Data

Download BraTS 2023 GLI data from Synapse (free registration required). Place volumes in data/raw/. See notebooks/01_data_exploration.ipynb for expected directory structure.

Training

python src/training/train.py --config configs/train_config.yaml

Inference

python src/inference/predict.py --input path/to/volume.nii.gz --output path/to/output/

Docker

docker build -f docker/Dockerfile -t brats-inference .
docker run -p 8000:8000 brats-inference

---

Dataset

BraTS 2023 GLI (Brain Tumor Segmentation Challenge — Glioma)

• Hosted via Synapse (RSNA-ASNR-MICCAI)
• ~1,200 multimodal MRI cases with expert annotations
• Four MRI modalities per case: T1, T1ce, T2, FLAIR
• Labels: 0 — background, 1 — necrotic core, 2 — peritumoral edema, 3 — enhancing tumor

---

Background

Developed as a portfolio project demonstrating full-stack ML engineering in clinical medical imaging. Informed by 8+ years of clinical experience in cardiac telemetry monitoring, with real-world awareness of clinical workflow constraints, alarm fatigue, and patient safety considerations that are often absent from purely academic implementations.

---

License

MIT License — see LICENSE for details.