| --- |
| license: other |
| license_name: research-only |
| license_link: LICENSE |
| tags: |
| - medical-imaging |
| - mri |
| - segmentation |
| - canine |
| - veterinary |
| - lora |
| - pytorch |
| - monai |
| library_name: pytorch |
| pipeline_tag: image-segmentation |
| --- |
| |
| # DeepCAN-SEG-PosEnc-T1 |
|
|
| **Canine Brain MRI 9-Class Segmentation — T1 sequence adapted** |
|
|
| A T1-weighted axial domain adaptation of [`hwonheo/DeepCAN-SEG-PosEnc`](https://huggingface.co/hwonheo/DeepCAN-SEG-PosEnc). |
| The shared base (T2-trained) was adapted to T1 with LoRA (rank 16, α 32) on Conv3d |
| layers; the adapters are **merged into the base weights**, so this checkpoint is a |
| standard `LRSegmentationMultiClassUNet` — a drop-in replacement loaded exactly 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: image intensity + normalized x/y/z position encoding (for L/R |
| hemisphere discrimination). Models expect the `RPS` orientation / 0.5 mm grid layout. |
|
|
| ## Performance (held-out T1 subjects, Dice, mean of classes 1–8) |
|
|
| | split | base (T2 model on T1) | **T1-adapted** | Δ | |
| | --- | --- | --- | --- | |
| | val | 0.322 | **0.649** | **+0.327** | |
| | test | 0.298 | **0.626** | **+0.328** | |
|
|
| Per-class val Dice: BG 0.96 · Cereb 0.79/0.81 · GM 0.69/0.69 · WM 0.58/0.61 · Vent 0.52/0.52. |
| The base T2 model nearly collapses on T1 (WM/Ventricle ~0.1–0.2); T1 adaptation roughly |
| doubles Dice with symmetric L/R recovery. |
|
|
| ## Training |
|
|
| | | | |
| | --- | --- | |
| | Base | DeepCAN-SEG-PosEnc (T2) | |
| | Method | LoRA (r=16, α=32) on all Conv3d, base frozen, adapters merged at export | |
| | Data | 30 T1 HR subjects → 1510 balanced 64³ patches @ 0.5 mm | |
| | Optimizer | AdamW, LR 2e-4, weight decay 1e-5 | |
| | Schedule | cosine, 200 epochs (early-stopped @ 40) | |
| | Loss | MultiClass Dice + CE (dice_weight 0.7) | |
| | W&B | https://wandb.ai/heohwon/DeepCAN-SegSR-public/runs/wrkvlfbo | |
| |
| ## Usage |
| |
| ```python |
| from huggingface_hub import snapshot_download |
| snapshot_download(repo_id="hwonheo/DeepCAN-SEG-PosEnc-T1", |
| local_dir="src/checkpoint/DeepCAN-SEG-PosEnc-T1") |
|
|
| # Loads exactly like the T2 base (plain UNet, in_channels auto-detected = 4) |
| from src.inference.models.segmentation_inferencer import SegmentationInferencer |
| seg = SegmentationInferencer( |
| checkpoint_path="src/checkpoint/DeepCAN-SEG-PosEnc-T1/DeepCAN-SEG-PosEnc-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](https://huggingface.co/hwonheo/DeepCAN-SEG-PosEnc-T1/resolve/main/LICENSE). |
| Contact: Hwon Heo, PhD (heohwon@gmail.com), BMC lab, Asan Medical Center. |
|
|
