A 12-month curriculum covering radiology, deep learning, nnU-Net, MONAI, and clinical deployment — everything you need to go from zero to competing in the BraTS Challenge at MICCAI.
Start the 12-Month PathA structured, month-by-month curriculum designed for students. Each month builds on the last — starting from clinical context, through preprocessing and model building, all the way to deployment and advanced radiomics.
You’re not training to be a radiologist — you’re training a model to behave like one. But you need to understand what the data means.
Magnetic Resonance Imaging (MRI) uses powerful magnets and radio waves to create detailed images of soft tissues — particularly the brain. Unlike CT scans or X-rays, MRI doesn’t use ionizing radiation, making it safer for repeated imaging. It produces images with excellent contrast between different types of brain tissue, ideal for spotting tumors.
An MRI scan is not a single 2D image — it’s a 3D volume made up of many 2D “slices.” Think of it like a loaf of bread: each slice is a cross-section of the brain. MRI files come in the NIfTI format (.nii or .nii.gz), which stores these 3D volumes along with metadata about voxel spacing and orientation.
We use four different “modalities” that each highlight different tissue properties — like looking at the same scene with different camera filters.
Brain tumors have distinct sub-regions. BraTS asks you to segment three nested regions:
From “what is a neural network” to confidently working with CNNs and U-Net.
Supervised learning: Learning from labeled examples (MRI + expert tumor boundaries).
CNNs: Convolution filters detect edges, textures, and patterns in images.
U-Net: Encoder-decoder with skip connections — the architecture behind nnU-Net.
Loss functions: Dice loss and cross-entropy — how the model measures segmentation mistakes.
Backpropagation: How the model learns through gradient descent.
Linear algebra: Matrices and multiplication (tensor operations). 3Blue1Brown’s series is the gold standard.
Basic calculus: Derivatives and the chain rule for backpropagation.
Probability: Softmax, sigmoid, distributions. Your model outputs per-voxel class probabilities.
nnU-Net automatically adapts to any dataset. It dominated the Medical Segmentation Decathlon and is the baseline for most BraTS competitors.
The name “no-new-Net” is the whole point — it squeezes maximum performance from the classic U-Net by optimizing everything else: preprocessing, augmentation, training schedule, patch size, network depth, post-processing, and ensembling. It analyzes your dataset’s “fingerprint” and auto-configures the pipeline.
It generates up to three configs: 2D U-Net, 3D full-resolution, and 3D cascade. It trains all with 5-fold cross-validation, then picks the best or ensembles them. Nine out of ten MICCAI 2020 winners built on nnU-Net.
Requirements: NVIDIA GPU (8GB+ VRAM), CUDA, PyTorch, Python 3.9+.
# Create environment
conda create -n nnunet python=3.10
conda activate nnunet
# Install PyTorch + nnU-Net
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install nnunetv2
# Set environment variables
export nnUNet_raw="/path/to/nnUNet_raw"
export nnUNet_preprocessed="/path/to/nnUNet_preprocessed"
export nnUNet_results="/path/to/nnUNet_results"nnU-Net v2 expects this folder structure:
Dataset001_BraTS/
├── imagesTr/
│ ├── BraTS_001_0000.nii.gz # T1
│ ├── BraTS_001_0001.nii.gz # T1ce
│ ├── BraTS_001_0002.nii.gz # T2
│ ├── BraTS_001_0003.nii.gz # FLAIR
├── labelsTr/
│ ├── BraTS_001.nii.gz
└── dataset.jsonThree commands:
# 1. Analyze & preprocess
nnUNetv2_plan_and_preprocess -d 001 --verify_dataset_integrity
# 2. Train (3D full-res, fold 0)
nnUNetv2_train 001 3d_fullres 0
# 3. Predict
nnUNetv2_predict -i INPUT -o OUTPUT -d 001 -c 3d_fullres -f 0Train all 5 folds for best results. Each fold: 12–48+ hours on a single GPU.
documentation/how_to_use_nnunet.md.Running since 2012, BraTS is the premier benchmark for brain tumor segmentation.
BraTS is an annual challenge at MICCAI, the top conference in medical image analysis. Organized by CBICA (UPenn) with RSNA, ASNR, ESNR, NIH, and FDA. Since 2023, it expanded into a “Cluster of Challenges” with 12+ tasks.
Task 1 — Adult Glioma: Pre- and post-treatment glioma segmentation.
Tasks 2–3 — Meningioma: Pre-treatment and pre-RT meningioma segmentation.
Task 4 — Brain Metastases: 1,475 cases, 4-label system, pre- and post-treatment.
Task 5 — BraTS-Africa: Addressing bias from underrepresentation of sub-Saharan populations.
Task 6 — Pediatric. Task 7 — GOAT (generalizability across tumor types).
Tasks 8–10: Synthesis, inpainting, and histopathology.
Apply your skills elsewhere and build your portfolio.
| Competition | Target | Modality | Details | Link |
|---|---|---|---|---|
| Medical Segmentation Decathlon | 10 organs/tumors | CT & MRI | Generalist benchmark. Rolling leaderboard still active. | medicaldecathlon.com |
| AMOS | 15 abdominal organs | CT & MRI | 500+ cases. Won by nnU-Net in 2022. | grand-challenge.org |
| KiTS | Kidneys & tumors | CT | 300+ cases. Simpler — good stepping stone. | kits-challenge.org |
| HECKTOR | Head & neck tumors | PET/CT & MRI | Multi-modal for radiation therapy planning. | grand-challenge.org |
| ISLES | Stroke lesions | MRI | Brain MRI like BraTS but for stroke. | isles-challenge.org |
| Grand-Challenge.org | Dozens active | Various | Largest biomedical challenge platform. | grand-challenge.org |
| Kaggle Medical Imaging | Various | Various | Beginner-friendly. Often has prize money. | kaggle.com |
Tools you’ll use throughout the curriculum and resources for going deeper.
nib.load('scan.nii.gz').get_fdata() gives you a NumPy array.