🧠 SegResNet — BraTS Brain Tumor Segmentation (3D, FP16)
This model fine-tunes SegResNet for 3D brain tumor segmentation (Tumor Core, Whole Tumor, Enhancing Tumor) using the
BraTS2020 Dataset.
The model was trained with mixed precision (FP16) using MONAI and Accelerate for efficient multi-GPU training.
⚙️ Configuration
| Attribute | Value |
|---|---|
| Base Model | SegResNet (MONAI) |
| Dataset | BraTS2020 (Kaggle) |
| Epochs | 150 |
| Batch Size (per GPU) | 2 |
| ROI Size | 128×128×128 |
| Optimizer | Adam (LR=0.0001, WD=1e-05) |
| Precision | FP16 (mixed=True) |
| Framework | MONAI + Accelerate |
| Device | Multi-GPU (num_processes=2) |
🩺 Example Segmentation
2D Prediction
3D Prediction
📈 Results (Final Test Metrics)
| Metric | Average | Tumor Core (TC) | Whole Tumor (WT) | Enhancing Tumor (ET) |
|---|---|---|---|---|
| Dice | 0.789 | 0.775 | 0.874 | 0.719 |
| Precision | 0.847 | 0.856 | 0.938 | 0.749 |
| Recall | 0.771 | 0.762 | 0.839 | 0.713 |
| Accuracy | 0.987 | 0.999 | 0.999 | 0.964 |
| Loss | 0.192 | - | - | - |
FP16 mixed precision enabled faster training while maintaining high segmentation accuracy.
No data augmentation was used to preserve original image fidelity.