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-SEG-PosEnc-T1-CanonAug

Canine Brain MRI 9-Class Segmentation — T1 + Canon-robust domain randomization

⚠️ Research / preview checkpoint. Trained with simulated Canon-style degradation on SKY (GE-family) T1 data — no real Canon-labeled data. See Known limitations before clinical routing.

A T1-weighted, Canon-robust adaptation of hwonheo/DeepCAN-SEG-PosEnc (shared T2-trained base). LoRA (rank 16, α 32) on Conv3d layers, trained with canon_aug intensity domain randomization (applied to the intensity channel only — the xyz position-encoding channels are never degraded). Adapters are merged into the base weights, so this is a standard LRSegmentationMultiClassUNet loaded exactly like the base model (drop-in).

Classes (9, L/R separated)

0 Background · 1/5 Lateral Ventricle L/R · 2/6 Gray Matter L/R · 3/7 White Matter L/R · 4/8 Cerebellum L/R

Input is 4-channel: image intensity + normalized x/y/z position encoding. Models expect the RPS orientation / 0.5 mm grid layout.

Domain randomization (`canon_aug`)

To mimic thick-slice, low-contrast Canon (Vantage Elan) clinical scans without any Canon labels, the intensity channel is degraded during training: anisotropic through-plane Gaussian blur (σ_z 1–3), GM-WM contrast gamma (0.6–1.6), mild bias field, and Gaussian noise. Validation uses a deterministic degradation (σ_z=2.0, γ=1.3) so val_dice and early-stopping track Canon-restoration.

Performance

val_dice 0.6565 on the Canon-degraded validation set (early-stopped @ epoch 74).

On the real Canon "달래" AX T1WI case, canon_aug restores near-normal symmetry vs the plain T1 model:

metric	plain T1	T1 + CanonAug
Hemisphere asymmetry	32.4%	13.0%
Cerebral-cortex asymmetry	45.3%	13.9%
White-matter asymmetry	20.3%	8.5%

Training


Base	DeepCAN-SEG-PosEnc (T2)
Method	LoRA (r=16, α=32) on all Conv3d + `canon_aug`, base frozen, adapters merged at export
Data	30 T1 HR subjects → balanced 64³ patches @ 0.5 mm
Optimizer	AdamW, LR 2e-4, weight decay 1e-5
Schedule	cosine, 200 epochs (early-stopped @ 74)
Loss	MultiClass Dice + CE (dice_weight 0.7)
W&B	https://wandb.ai/heohwon/DeepCAN-SegSR-public/runs/la5fb6qx

Usage

from huggingface_hub import snapshot_download
snapshot_download(repo_id="hwonheo/DeepCAN-SEG-PosEnc-T1-CA",
                  local_dir="src/checkpoint/DeepCAN-SEG-PosEnc-T1-CanonAug-LoRA")

from src.inference.models.segmentation_inferencer import SegmentationInferencer
seg = SegmentationInferencer(
    checkpoint_path="src/checkpoint/DeepCAN-SEG-PosEnc-T1-CanonAug-LoRA/DeepCAN-SEG-PosEnc-T1-CanonAug-LoRA.pth",
    device="cuda")

Known limitations

Trained on GE-family (SKY) T1 with simulated Canon degradation, not real Canon anatomy — domain randomization narrows but does not fully close the gap.
Smaller T1 corpus (30 subjects) than the T2 base; Ventricle is the weakest class.

License

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

Downloads last month: -