You can deploy this model and start earning money today!

MedOtter/
medotter-glioma-attunet

Name: medotter-glioma-attunet
Availability: InStock

Views

No views yet

MedOtter Glioma — Attention U-Net (WT/TC/ET)

A 2D Attention U-Net for autonomous (prompt-free) glioma sub-region segmentation, submitted to the MedOtter leaderboard glioma_wt_tc_et task.

Architecture: Attention U-Net (~34.9M params), inlined in predictor.py.
Training data: BraTS2023-GLI (t1c modality), 2D axial slices.
Input at inference: the worker passes a 4-channel volume (t1n, t1c, t2w, t2f); the predictor uses t1c only.
Output: {0=background, 1=NCR, 2=edema, 3=enhancing} on the native voxel grid.

How it runs

predictor.py is self-contained — it defines the model inline and pulls weights.pth from this repo at load() time. The evaluation worker only needs predictor.py. Per axial slice: brain bounding-box crop → percentile(0.5/99.5) uint8 → replicate to 3 channels → resize 256 → ImageNet normalize → argmax → map back onto the native grid.

Single-modality (t1c) baseline; WT/edema are harder than TC/ET because they are best seen on FLAIR/T2, which this baseline does not use.

MedOtter/medotter-glioma-attunet

MedOtter Glioma — Attention U-Net (WT/TC/ET)

How it runs

MedOtter/
medotter-glioma-attunet