DeepCAN-SR-swinViT-T1

Canine Brain MRI Super-Resolution (SwinUNETR-SR) — T1 sequence adapted

A T1-weighted axial domain adaptation of hwonheo/DeepCAN-SR-swinViT. The shared SwinUNETR-SR base was adapted to T1 with LoRA (rank 32, α 64) on the swinViT encoder while training the decoder + output head; the swinViT backbone stays frozen. The checkpoint keeps its LoRA structure and loads into SwinUNETRSR(use_lora=True) exactly like the base model (drop-in).

Performance (held-out T1 subjects)

split	metric	LR input (floor)	base (T2 model on T1)	T1-adapted	Δ
val	PSNR	28.1 dB	32.98 dB	35.70 dB	+2.72 dB
val	SSIM	—	0.917	0.961	+0.044
test	PSNR	27.7 dB	32.53 dB	35.10 dB	+2.57 dB
test	SSIM	—	0.930	0.958	+0.028

Training

Base	DeepCAN-SR-swinViT (T2)
Method	LoRA (r=32, α=64) on swinViT; encoder frozen, decoder + out + adapters trained
Data	30 T1 HR subjects → 64³ LR→HR pairs @ 0.5 mm (Z-blur + Rician LR simulation)
Optimizer	AdamW, LR 5e-5, weight decay 1e-5
Schedule	cosine, 100 epochs (early-stopped @ 39)
Loss	Combined L1 + SSIM (0.1) + gradient (0.05)
W&B	https://wandb.ai/heohwon/DeepCAN-SegSR-public/runs/dktrk6x4

Usage

from huggingface_hub import snapshot_download
snapshot_download(repo_id="hwonheo/DeepCAN-SR-swinViT-T1",
                  local_dir="src/checkpoint/DeepCAN-SR-swinViT-T1")

from src.inference.models.sr_inferencer import SRInferencer
sr = SRInferencer(
    checkpoint_path="src/checkpoint/DeepCAN-SR-swinViT-T1/DeepCAN-SR-swinViT-T1.pth",
    device="cuda")

In the clinical pipeline, T1 axial scans are auto-detected from DICOM metadata (EchoTime/RepetitionTime/SeriesDescription) and routed to this checkpoint.

License

Research use only — see LICENSE. Contact: Hwon Heo, PhD (heohwon@gmail.com), BMC lab, Asan Medical Center.