Evaluation Results
Trained on the full Kaggle Brain Tumor 12K dataset — 8,673 train / 1,858 val / 1,860 test images. All modalities loaded: T1, T1C+, T2.
| Parameter | Value |
|---|---|
| Model | 3D Attention U-Net |
| Parameters | 2.2M |
| Base filters | 32 |
| Depth | 2 |
| Image size | 64×64 |
| D-frames (pseudo-3D) | 2 |
| Epochs | 50 |
| Batch size | 4 |
| Optimizer | Adam, lr=1e-3 |
| Scheduler | ReduceLROnPlateau (factor=0.5, patience=5) |
| Loss | HybridLoss (α=0.5 Dice, β=0.3 Focal, γ=0.2 Boundary) |
| Classes | 8 (background + 7 WHO tumor categories) |
| Dataset split | 70% train / 15% val / 15% test |
| Hardware | CPU (~43 min/epoch, ~36 hours total) |
| Class | Val Dice (epoch 50) | Test Dice (1,860 images) |
|---|---|---|
| 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 |
Glioma scores lowest of the tumor classes — expected, since gliomas are the most morphologically heterogeneous category (varying grade, shape, infiltration pattern), while Germ Cell and Nerve Sheath tumors tend to be more consistent and well-circumscribed on MRI.
Val and test scores track closely across all classes, indicating the model generalizes rather than memorizing the validation set.
Val Mean Tumor Dice climbed from 0.63 → 0.74 across 50 epochs, train/val gap stayed small throughout — no significant overfitting.
See results/ in the repo for inference on all 7 tumor categories.
The best checkpoint (best_model_dice_0.7350.pt) is attached to the GitHub Release v2.0.0.
Download and run inference directly without training from scratch:
# Run inference on a real image
python scripts/predict3d.py \
--checkpoint checkpoints/best_model_dice_0.7350.pt \
--image data/raw/Images_/Gliomas/T1C+/subtype/image.jpg \
--out prediction.png