surface_m7_nnunet

An nnU-Net (v2) model for surface segmentation of Herculaneum papyrus CT volumes. This is the nnU-Net component (internally "m7") of the 1st-place solution to the Kaggle Vesuvius Challenge – Surface Detection competition.

Note: This is a partial component of that solution — the standalone nnU-Net model — not the full ensemble/architecture described in the writeup.

Source writeup: 1st-place solution for the Vesuvius Challenge – Surface Detection

Model details

• Framework: nnU-Net v2
• Dataset: Dataset100_VesuviusSurface (786 training volumes)
• Input: single channel CT, read from .tif via nnU-Net's SimpleTiffIO
• Labels: background = 0, surface = 1, ignore = 2
• Plans: nnUNetResEncUNetLPlans — Residual Encoder U-Net, "L" preset (dynamic_network_architectures...ResidualEncoderUNet)
• Configurations defined in the plans:
  • 2d — patch size 320 × 320
  • 3d_fullres — patch size 192 × 192 × 192, spacing 1.0³
• Checkpoint: best checkpoint for a single fold (fold_0).

Files

dataset.json                 # channels / labels / dataset metadata
dataset_fingerprint.json     # nnU-Net dataset fingerprint
plans.json                   # nnU-Net plans (2d + 3d_fullres configs)
fold_0/
  checkpoint_best.pth        # trained weights (best checkpoint), ~783 MB

The layout is the standard nnU-Net trained-model folder, so it can be used directly as an nnUNet_results model directory.

Usage

Download the repo and point nnU-Net v2 at it as a results directory:

from huggingface_hub import snapshot_download

model_dir = snapshot_download(repo_id="scrollprize/surface_m7_nnunet")
# `model_dir` now contains dataset.json, plans.json, fold_0/checkpoint_best.pth

Then run inference with nnUNetv2_predict (or the nnUNetPredictor API), selecting the configuration that matches this checkpoint (3d_fullres or 2d) and -f 0 for the single provided fold. See the nnU-Net v2 inference docs.

Attribution & license

The model originates from the 1st-place Kaggle Vesuvius Challenge – Surface Detection solution (linked above); please credit the original authors. Released here under Apache-2.0 to match the other Scroll Prize surface models. If the original authors specify different terms, those govern.