You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

DeepCAN-SR-swinViT: Canine Brain MRI Super-Resolution Model

GitHub: https://github.com/Core-BMC/DeepCAN-SegSR.git

4x super-resolution model for canine brain MRI using SwinUNETR with LoRA adapters, fine-tuned on v1.1a high-resolution balanced dataset.

Model Description

This model performs 2x super-resolution on canine brain MRI scans, enhancing low-resolution inputs to high-resolution outputs for improved visualization and downstream analysis.

Architecture: SwinUNETR-SR with LoRA fine-tuning
Input: 64x64x64 patches (single channel)
Output: 64x64x64 patches (2x enhanced resolution)
Parameters: ~731MB (includes pretrained SwinUNETR backbone)
LoRA Config: r=32, alpha=64, dropout=0.1

Super-Resolution Pipeline

Input (0.5mm) → SwinUNETR-SR → Output (0.25mm equivalent)

Performance

Validation Metrics (Epoch 100)

Metric	Value
PSNR	35.88 dB
SSIM	0.972
Loss	0.010

Training Metrics

Train PSNR: 37.36 dB
Train Loss: 0.009

Comparison with Previous Version

Metric	v1.0 (50 epochs)	v1.1a (100 epochs)	Improvement
Val PSNR	32.23 dB	35.88 dB	+3.65 dB
Val SSIM	0.941	0.972	+0.031
Val Loss	0.016	0.010	-37.5%

Training Details

Dataset: DeepCAN v1.1a (balanced L/R hemisphere patches)
Base Model: DeepCAN-SR-swinViT v1.0 (fine-tuned)
Fine-tuning: LoRA on swinViT layers
Epochs: 100
Batch Size: 4
Learning Rate: 5e-5 (cosine scheduler, min_lr=1e-6)
Optimizer: AdamW (weight_decay=1e-5)
Loss: Combined L1 + SSIM + Gradient (weights: 1.0, 0.1, 0.05)
Early Stopping: patience=15, monitor=val_psnr
Hardware: NVIDIA RTX 4090 (24GB)
Training Time: ~72 hours

Training Logs

Full training logs available on Weights & Biases:

Project: DeepCAN-SegSR
Run: DeepCAN-SR-swinViT

Usage

import torch
from monai.networks.nets import SwinUNETR

# Load model
model = SwinUNETR(
    img_size=(64, 64, 64),
    in_channels=1,
    out_channels=1,
    feature_size=48,
    use_checkpoint=False
)
checkpoint = torch.load("DeepCAN-SR-swinViT.pth", map_location="cpu")
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# Inference
with torch.no_grad():
    # input_patch: [B, 1, 64, 64, 64] - normalized intensity
    output = model(input_patch)

With Clinical Pipeline

# Clone the main repository
git clone https://github.com/Core-BMC/DeepCAN-SegSR.git
cd DeepCAN-SegSR

# Run super-resolution only
python -m src.inference.cli sr \
    --input data/lr_volume.nii.gz \
    --checkpoint path/to/DeepCAN-SR-swinViT.pth

# Or use full clinical pipeline (SR + Segmentation)
python -m src.inference.cli clinical \
    --input your_dicom_folder/ \
    --output outputs/

Model Files

DeepCAN-SR-swinViT.pth: Model weights (731MB)

Requirements

PyTorch 2.0+
MONAI 1.3+
CUDA 11.8+ (for GPU inference)

Limitations

Trained on canine brain MRI only (not validated for other species)
Optimized for T2-weighted sequences
Requires preprocessing to match training data distribution
Research use only - not validated for clinical diagnosis

Citation

@software{deepcan2025,
  title = {DeepCAN SegSR Suite: Canine Brain MRI Super-Resolution and Segmentation},
  author = {Hwon Heo & Woo Hyun Shim, DeepCAN AI team},
  year = {2025},
  url = {https://github.com/Core-BMC/DeepCAN-SegSR}
}

License

This model is released under the DeepCAN Research License - free for non-commercial research and educational use only.

For commercial licensing inquiries, contact: heohwon@gmail.com

See LICENSE for full terms.

Downloads last month: -; Downloads are not tracked for this model. How to track