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-T2-CanonAug

Canine Brain MRI 9-Class Segmentation — T2 + Canon-robust domain randomization

Model version 20260702-SEG-T2-CA — full-dataset (858-subject) LoRA. Supersedes the 150-subject pilot; the pilot is retained locally under _legacy/.

A Canon-robust adaptation of hwonheo/DeepCAN-SEG-PosEnc. LoRA (rank 16, α 32) on Conv3d layers, trained with canon_aug intensity domain randomization (intensity channel only; xyz position-encoding never degraded). Adapters are merged into the base weights — a standard LRSegmentationMultiClassUNet, drop-in loadable like the base model.

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 (intensity + normalized xyz position encoding), RPS orientation / 0.5 mm grid.

Domain randomization (`canon_aug`)

Anisotropic through-plane blur (σ_z 1–3), GM-WM contrast gamma (0.6–1.6), mild bias field, and noise applied to the intensity channel to mimic Canon thick-slice scans (position channels untouched).

Performance

Held-out T2 validation split (plain patches, 9-class mean): val_dice 0.8446 (early-stopped @ epoch 22 of 120). Per-class dice is L/R balanced — the right-hemisphere training deficit of the pilot is resolved:

	Vent	GM	WM	Cereb
Left	0.820	0.846	0.791	0.906
Right	0.832	0.845	0.809	0.908

Real Canon "달래" AX T2WI (Vantage Elan, 2.8 mm)

The full model corrects the pilot's severe left-biased leakage into an anatomically balanced segmentation (same clinical pipeline, |asymmetry| %):

metric	pilot (150)	full (858)
Hemisphere asymmetry	37.98%	5.68%
Ventricle asymmetry	58.74%	6.22%
Cortex asymmetry	48.16%	13.08%
Cerebellum asymmetry	56.01%	32.03%
White-matter asymmetry	5.40%	10.66%
Total brain volume	70.82 mL	54.13 mL

The pilot's 70.8 mL sat "in range" only because a leaky, left-dominant over-segmentation inflated it; the full model's 54.1 mL reflects a clean, symmetric mask (it trips the >60 mL QC heuristic — that threshold is calibrated on larger reference scans and is being revisited). On a normal GE SIGNA Explorer T2 case the full model shows no regression (62.3 → 63.1 mL, L/R ratio 1.02).

Training


Base	DeepCAN-SEG-PosEnc (T2)
Method	LoRA (r=16, α=32) on all Conv3d + `canon_aug`, base frozen, adapters merged at export
Data	858 T2 HR subjects (full) → 64³ patches @ 0.5 mm
Optimizer	AdamW, LR 2e-4, weight decay 1e-5
Schedule	cosine, 120 epochs (early-stopped @ 22)
Loss	MultiClass Dice + CE (dice_weight 0.7), gradual sqrt class weights
W&B	https://wandb.ai/heohwon/DeepCAN-SegSR-public/runs/82xpwoqc

Usage

from huggingface_hub import snapshot_download
snapshot_download(repo_id="hwonheo/DeepCAN-SEG-PosEnc-T2-CA",
                  local_dir="src/checkpoint/DeepCAN-SEG-PosEnc-T2-CanonAug-LoRA")

from src.inference.models.segmentation_inferencer import SegmentationInferencer
seg = SegmentationInferencer(
    checkpoint_path="src/checkpoint/DeepCAN-SEG-PosEnc-T2-CanonAug-LoRA/DeepCAN-SEG-PosEnc-T2-CanonAug-LoRA.pth",
    device="cuda")
print(seg.model_version)  # 20260702-SEG-T2-CA

Known limitations

Canon thick-slice (2.8–3 mm) cerebellum asymmetry (~32%) and mild white-matter right-bias remain; through-plane resolution still limits GM-WM separability.
Total-brain QC threshold (>60 mL) is calibrated on higher-resolution reference scans and can false-warn on clean Canon segmentations — under review.
Trained on simulated Canon degradation with no real Canon-labeled data; real Canon labels are the planned next lever.

License

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

Downloads last month: -