v2.0.0 — 8-Class WHO Tumor Segmentation (Test Dice 0.7387)

Brain Tumor Segmentation v2.0.0 — 8-Class WHO Segmentation

Major upgrade from binary tumor/background segmentation to full 8-class WHO tumor category segmentation, trained on the complete 12K dataset across all MRI modalities.

What's new

• 8-class output: Background + 7 WHO tumor categories (Glioma, Meningioma, Nerve Sheath, Embryonic, Mixed Neuronal, Mesenchymal, Germ Cell)
• Full dataset: All 12,391 images across T1, T1C+, and T2 modalities (previously T1c-only, 5,110 images)
• Weakly-supervised pseudo-labels: Binary consensus masks + folder-level WHO category → per-pixel multi-class labels
• Bigger model: 2.2M parameters (base_filters 16→32), 4× larger than v1.0.0
• Resume support: Full optimizer + scheduler state checkpointing
• Automatic test-set evaluation: Runs after training on the 1,860-image held-out test split

Results (50 epochs, CPU)

Class	Val Dice	Test Dice
Mean Tumor Dice	0.7350	0.7387
Background	0.9918	0.9917
Glioma	0.4671	0.4713
Meningioma	0.6659	0.6927
Nerve Sheath	0.7878	0.8077
Embryonic	0.7703	0.7535
Mixed Neuronal	0.7247	0.7391
Mesenchymal	0.8273	0.7786
Germ Cell	0.9020	0.9278

Dataset split: 8,673 train / 1,858 val / 1,860 test (70/15/15).

Checkpoint

best_model_dice_0.7350.pt — 3D Attention U-Net, base_filters=32, depth=2, image_size=64, d_frames=2.

wget https://github.com/motazalqaoud/Brain-Tumor-Segmentation/releases/download/v2.0.0/best_model_dice_0.7350.pt \
     -O checkpoints/best_model_dice_0.7350.pt

python scripts/predict3d.py \
  --checkpoint checkpoints/best_model_dice_0.7350.pt \
  --image path/to/brain_mri.jpg \
  --out prediction.png

See the Evaluation Results wiki page for full training curves and per-class analysis.

v1.0.0 — Brain Tumor Segmentation (Val Dice 0.7894)

First release of Brain Tumor Segmentation — 3D Attention U-Net

Trained for 50 epochs on the Kaggle Brain Tumor 12K dataset using a CPU-optimised pseudo-3D pipeline.

Results

Metric	Value
Val Dice	0.7894
Train Dice	0.9000
Overfitting gap	0.11 ✓
Model parameters	559K
Training time	~6 hours (CPU)

Checkpoint

best_model_dice_0.7894.pt (6.6 MB) is attached — download and run inference without retraining:

python scripts/predict3d.py \
  --checkpoint best_model_dice_0.7894.pt \
  --image path/to/brain_mri.jpg \
  --out prediction.png

What's included

• 3D Attention U-Net with SE channel attention and spatial attention gates
• Hybrid loss (Weighted Dice + Focal + Boundary)
• Hardware presets: CPU / 8GB GPU / 16GB GPU
• run.py automation for non-technical users
• Full inference pipeline with confidence maps

See the full README and Evaluation Results wiki page for details.