emiraran/vesuvius-3d-surface-segmentation

Vesuvius Challenge — 3D Papyrus Surface Segmentation Ensemble

10-model multi-architecture PyTorch 3D ensemble (5x PureSegResNet3D + 5x ResEncUNet3D) trained for 3D volumetric surface segmentation of carbonized Herculaneum papyrus scrolls in micro-CT scans (Vesuvius Challenge - Surface Detection).

• GitHub Source Code (Training, Metrics, Inference & Topological Post-Processing): EmirMuhammetARAN/vesuvius-3d-surface-segmentation
• Kaggle Competition Private Score: 0.54788 (Public: 0.54030) — improved from 0.40420 v1 baseline (+0.14368)

---

Included Ensemble Checkpoints (weights/)

#	Checkpoint File	Architecture	Params	Patch Size	Split / Run	Weight	Size
1	weights/segresnet3d_p192_finetune192_best.pth	PureSegResNet3D	18,795,969	192^3	All-Data fine-tune best (L4)	4.5	71.73 MB
2	weights/segresnet3d_p192_finetune192_epoch150.pth	PureSegResNet3D	18,795,969	192^3	All-Data fine-tune Ep 150 snap	2.5	71.74 MB
3	weights/segresnet3d_p160_fold2_best.pth	PureSegResNet3D	18,795,969	160^3	Fold 2 best	3.5	71.73 MB
4	weights/segresnet3d_p160_best.pth	PureSegResNet3D	18,795,969	160^3	Fold 1 best	3.0	71.73 MB
5	weights/segresnet3d_p192_colab_fold3_best.pth	PureSegResNet3D	18,795,969	192^3	Fold 3 best	2.5	71.73 MB
6	weights/resenc_unet3d_p192_alldata_best.pth	ResEncUNet3D	15,912,097	192^3	All-Data best	2.0	60.73 MB
7	weights/resenc_unet3d_p192_colab_fold3_best.pth	ResEncUNet3D	15,912,097	192^3	Fold 3 best	2.0	60.74 MB
8	weights/resenc_unet3d_p192_v2_best.pth	ResEncUNet3D	15,912,097	192^3	Fold 1 best	1.5	60.73 MB
9	weights/resenc_unet3d_p192_fold2_best.pth	ResEncUNet3D	15,912,097	192^3	Fold 2 best	1.0	60.73 MB
10	weights/resenc_unet3d_p160_p160_best.pth	ResEncUNet3D	15,912,097	160^3	Fold 1 best	1.0	60.73 MB

---

Competition Progression

Stage	Public LB	Private LB
v1 TensorFlow 3D U-Net (96^3)	0.39477	0.40420
v2 Initial PyTorch (128^3)	0.46164	0.47161
v2 Progressive Scale (160^3/192^3)	0.52229	0.52719
v2 8-Model Ensemble	0.52601	0.53265
v2 10-Model Ensemble (192^3 All-Data Fine-Tune)	0.54030	0.54788

---

Usage (Standalone PyTorch — Zero MONAI Dependency)

import torch
from huggingface_hub import hf_hub_download
from importlib.util import spec_from_file_location, module_from_spec

modeling_path = hf_hub_download("emiraran/vesuvius-3d-surface-segmentation", "modeling.py")
spec = spec_from_file_location("vesuvius_modeling", modeling_path)
mod = module_from_spec(spec)
spec.loader.exec_module(mod)

ckpt_path = hf_hub_download(
    "emiraran/vesuvius-3d-surface-segmentation",
    "weights/segresnet3d_p192_finetune192_best.pth",
)
model = mod.load_model("PureSegResNet3D", ckpt_path, device="cuda" if torch.cuda.is_available() else "cpu")

# Example 3D normalized CT patch: (B, 1, D, H, W) in [0, 1]
x = torch.zeros(1, 1, 192, 192, 192)
with torch.no_grad():
    probs = torch.sigmoid(model(x))

Full sliding-window TTA inference, training, and topological post-processing code are available in the GitHub repository.

License

MIT License