README.md

metadata

title: BraTS Segmentation
emoji: 🧠
colorFrom: blue
colorTo: green
sdk: docker
pinned: false

Brain Tumor Segmentation with BraTS2020

A full-stack 3D U-Net implementation for brain tumor segmentation.

Stack: Python · NumPy · PyTorch · nibabel · FastAPI · React
Dataset: BraTS2020 — 369 training cases, 125 validation cases

---

Dataset Structure

MICCAI_BraTS2020_TrainingData/
  BraTS20_Training_001/
    BraTS20_Training_001_flair.nii   ← FLAIR modality
    BraTS20_Training_001_t1.nii      ← T1-weighted
    BraTS20_Training_001_t1ce.nii    ← T1 contrast-enhanced
    BraTS20_Training_001_t2.nii      ← T2-weighted
    BraTS20_Training_001_seg.nii     ← Ground truth mask
  BraTS20_Training_002/ ...
  BraTS20_Training_369/
  name_mapping.csv
  survival_info.csv

Each volume: shape (240, 240, 155), voxel size 1mm³, dtype float64.

---

Check out my beginner-friendly explanation for the dataset: Understanding and Processing the Brain Tumor Segmentation (BraTS2020) Dataset

Key Data Facts

Property	Value	Implication
Volume shape	(240, 240, 155)	3D — needs 3D-aware model
Voxel size	1mm × 1mm × 1mm	Physical scale known
Non-zero fraction	~15%	Crop before feeding to model
Intensity ranges	T2: 0–376, T1ce: 0–1845	Normalize per modality
Seg labels	{0, 1, 2, 4}	Remap 4 → 3 before training
Background fraction	97.63%	Cross-entropy alone won't work

MRI Modalities

Modality	Measures	Key use
T1	Longitudinal relaxation	Anatomy, grey/white matter
T1ce	T1 + gadolinium contrast	Active tumor (enhancing)
T2	Transverse relaxation	Fluid, edema
FLAIR	T2 with free water suppressed	Peritumoral edema

BraTS Evaluation Regions

Region	Labels	Clinical meaning
Whole Tumor (WT)	{1, 2, 3}	Total tumor extent
Tumor Core (TC)	{1, 3}	Surgically targetable core
Enhancing Tumor (ET)	{3}	Active, contrast-enhancing

---

Preprocessing Pipeline

Raw NIfTI (240×240×155)
        ↓
normalize_modality()     Z-score within brain mask, per modality
        ↓
crop_to_brain()          Tight bounding box of non-zero voxels
        ↓
resize_volume()          Trilinear interpolation → (128, 128, 128)
        ↓
Model input tensor       (4, 128, 128, 128) — 4 modalities stacked

normalize_modality

def normalize_modality(vol: np.ndarray) -> np.ndarray:
    brain_mask = vol > 0
    if brain_mask.sum() == 0:
        return vol
    mu  = vol[brain_mask].mean()
    std = vol[brain_mask].std() + 1e-8
    normalized = np.zeros_like(vol)
    normalized[brain_mask] = (vol[brain_mask] - mu) / std
    return normalized.astype(np.float32)

Why: MRI intensities are scanner-dependent — not comparable across patients or modalities. Z-score within brain voxels gives each modality mean≈0, std≈1, regardless of scanner.

crop_to_brain

def crop_to_brain(vol: np.ndarray) -> np.ndarray:
    coords = np.array(np.where(vol > 0))
    if coords.shape[1] == 0:
        return vol
    mins = coords.min(axis=1)
    maxs = coords.max(axis=1) + 1
    return vol[mins[0]:maxs[0], mins[1]:maxs[1], mins[2]:maxs[2]]

Why: 85% of the (240,240,155) volume is background air. Cropping to the brain bounding box removes wasted computation and GPU memory.

resize_volume

def resize_volume(vol: np.ndarray, target=(128, 128, 128)) -> np.ndarray:
    tensor  = torch.from_numpy(vol).float().unsqueeze(0).unsqueeze(0)
    resized = F.interpolate(tensor, size=target, mode="trilinear", align_corners=True)
    return resized.squeeze().numpy()

Why: The model requires fixed-size input. 128³ fits in ~10GB VRAM at batch size 1. align_corners=True ensures image and mask stay spatially aligned when resized separately.

---

Segmentation Label Remapping

BraTS2020 raw labels: {0, 1, 2, 4} — no label 3 (historical artifact).
Before training, remap 4 → 3 so model output indices match:

seg[seg == 4] = 3
# 0 = background
# 1 = necrotic core (NCR)
# 2 = peritumoral edema (ED)
# 3 = enhancing tumor (ET)  ← was label 4 in raw file

---

Repository Structure

Brain-Tumor-Segmentation-with-BraTS/
├── README.md
├── src/
│   ├── dataset.py               ← BraTSDataset             
│   ├── model.py                 ← 3D U-Net                 
│   ├── train.py                 ← training loop            
│   └── inference.py             ← inference + export              
└── frontend/                    ← React UI                 

---

References

• BraTS 2020 Challenge
• 3D U-Net (Çiçek et al., 2016)
• nnU-Net (Isensee et al., 2021)
• Dice Loss for medical segmentation

---