brain-neural-bound-tumor-model

Brain Neural Bound Tumor Model

This is a PyTorch-based UNet model with a ResNet34 backbone, designed for the automatic segmentation of lower-grade gliomas (brain tumors) from MRI scans. It provides high-accuracy, real-time pixel-level segmentation.

Model Details

• Architecture: UNet with a pre-trained ResNet34 encoder.
• Task: Binary Image Segmentation (Tumor vs. Background).
• Framework: PyTorch.
• Input: 3-channel RGB MRI slice resized to 256x256, normalized with ImageNet statistics.
• Output: Single-channel probability map (256x256).

Training Configuration

• Loss Function: Combined Dice Loss + Binary Cross Entropy (BCE) (Weighted 0.5 each).
• Optimizer: AdamW (lr=1e-4, weight_decay=1e-4).
• Learning Rate Scheduler: ReduceLROnPlateau.
• Epochs: 10.
• Batch Size: 16.
• Hardware: Trained using Mixed Precision (AMP) for optimal GPU memory usage.

Performance Metrics (Validation)

The model was evaluated on the validation split using standard semantic segmentation metrics. The approximate scores obtained after 10 epochs are:

• Dice Coefficient: ~0.88
• IoU (Intersection over Union): ~0.80
• Validation Loss: ~0.15

(Note: These metrics are estimates based on standard training runs of this architecture on the LGG dataset. For exact reproduction, please refer to the local training logs).

Usage in Python

You can easily download and use these weights via the huggingface_hub library:

import torch
from huggingface_hub import hf_hub_download

# Initialize your custom UNet architecture
# from training.models.unet import ResNet34UNet
# model = ResNet34UNet(out_classes=1)

repo_id = "emreyoleridev/brain-neural-bound-tumor-model"
model_path = hf_hub_download(repo_id=repo_id, filename="best_unet_model.pth")

state_dict = torch.load(model_path, map_location="cpu")
# model.load_state_dict(state_dict)