DeepCAN-SR-swinViT-T2-CanonAug
Canine Brain MRI Super-Resolution (SwinUNETR-SR) — T2 + Canon-robust domain randomization
⚠️ Research / preview checkpoint. Pilot-scale (150-subject) LoRA trained with simulated Canon degradation — no real Canon data.
A Canon-robust adaptation of hwonheo/DeepCAN-SR-swinViT.
LoRA (rank 32, α 64) on the swinViT encoder (backbone frozen, decoder + output
head trained), plus canon_aug LR-only domain randomization so the model
restores texture / contrast from Canon-like thick-slice input while the HR target
stays clean. The checkpoint keeps its LoRA structure and loads into
SwinUNETRSR(use_lora=True) exactly like the base model (drop-in).
Domain randomization (canon_aug)
The LR input only is degraded during training — anisotropic through-plane blur
(σ_z 1–3), contrast gamma (0.6–1.6), mild bias field, Gaussian noise — while the HR
target is untouched. Validation uses deterministic degradation (σ_z=2.0, γ=1.3) so
val_psnr measures Canon-restoration and early-stopping stays aligned with the goal.
Performance
On the Canon-degraded validation set: PSNR 33.08 dB · SSIM 0.943 (epoch 40).
Training
| Base | DeepCAN-SR-swinViT (T2) |
| Method | LoRA (r=32, α=64) on swinViT + canon_aug; encoder frozen, decoder + out + adapters trained |
| Data | 150 T2 HR subjects (pilot) → 64³ LR→HR pairs @ 0.5 mm |
| Optimizer | AdamW, LR 5e-5, weight decay 1e-5 |
| Schedule | cosine, 40 epochs |
| Loss | Combined L1 + SSIM (0.1) + gradient (0.05) |
| W&B | https://wandb.ai/heohwon/DeepCAN-SegSR-public/runs/ezl7ngdi |
Usage
from huggingface_hub import snapshot_download
snapshot_download(repo_id="hwonheo/DeepCAN-SR-swinViT-T2-CA",
local_dir="src/checkpoint/DeepCAN-SR-swinViT-T2-CanonAug-LoRA")
from src.inference.models.sr_inferencer import SRInferencer
sr = SRInferencer(
checkpoint_path="src/checkpoint/DeepCAN-SR-swinViT-T2-CanonAug-LoRA/DeepCAN-SR-swinViT-T2-CanonAug-LoRA.pth",
device="cuda")
Known limitations
- Trained on GE-family (SKY) T2 with simulated Canon degradation, not real Canon scans; pilot scale (150 of 858 available subjects).
License
Research use only — see LICENSE. Contact: Hwon Heo, PhD (heohwon@gmail.com), BMC lab, Asan Medical Center.
- Downloads last month
- -