Vesuvius Challenge — 3D Surface Detection

Deep learning solution for detecting ancient papyrus surfaces from 3D X-ray CT scans of carbonized Herculaneum scrolls. Built with PyTorch, MONAI, and trained on A100 GPUs via Modal.

Competition: Kaggle — Vesuvius Challenge Surface Detection

The Problem

Carbonized papyrus scrolls from Herculaneum (buried by Mt. Vesuvius in 79 AD) are too fragile to unroll. High-resolution 3D X-ray CT scans can reveal the internal structure — but detecting the actual papyrus surfaces within the volume is a 3D segmentation challenge requiring topology-aware models.

Architecture

Four model architectures, all operating on 3D volumetric data:

Model	Approach	Key Feature
3D U-Net	Symmetric encoder-decoder	Skip connections for fine detail
Attention U-Net	Gated attention gates	Focuses on salient regions
V-Net	Residual volumetric blocks	Dice-based training objective
Double U-Net	Stacked architecture	Two-stage refinement

Final predictions use ensemble + test-time augmentation (flips, rotations) for robust surface detection.

Training Pipeline

Loss Function

Custom TopologyAwareLoss combining three components:

loss = α · DiceLoss + β · FocalLoss + γ · BoundaryLoss
# α=0.5, β=0.3, γ=0.2

• Dice Loss — Overall segmentation quality
• Focal Loss — Hard example mining (γ=2.0)
• Boundary Loss — Surface edge accuracy (topology-aware)

Key Discovery

This is a 3D volumetric segmentation task, not 2D surface detection:

Cross-section of papyrus scroll:
    ╱╲╱╲╱╲╱╲╱╲╱╲    ← Multiple wrapped layers
   ╱  ╲  ╱  ╲  ╱  ╲   
  ╱    ╲╱    ╲╱    ╲  ← Each layer = papyrus surface
 ╱                  ╲

In a 3D chunk, papyrus can occupy 40–96% of voxels — high coverage predictions are correct, not corrupted.

Quick Start

Local Setup

git clone https://github.com/Jonathan-321/vesuviusChallenge.git
cd vesuviusChallenge

python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# Download competition data
kaggle competitions download -c vesuvius-challenge-surface-detection
unzip vesuvius-challenge-surface-detection.zip -d data/raw/

python scripts/preprocessing/prepare_data.py

Train on Modal (A100)

bash scripts/upload_to_modal.sh        # Upload data (one time)
modal run modal_training.py --command train     # Single model
modal run modal_training.py --command parallel  # Multi-model

Project Structure

vesuviusChallenge/
├── vesuvius-competition/
│   ├── src/
│   │   ├── training/        # Trainer, losses (TopologyAwareLoss)
│   │   ├── inference/       # Prediction + submission generation
│   │   └── evaluation/      # Metrics (Dice, IoU)
│   ├── configs/
│   │   ├── base/            # Default hyperparameters
│   │   └── experiments/     # Per-model configs
│   ├── scripts/             # Preprocessing, validation sweeps
│   └── notebooks/           # Kaggle inference notebooks
├── scripts/cluster/         # Cloud training orchestration
└── images/                  # README visuals

Tech Stack

Component	Tool
Framework	PyTorch + MONAI
Training	Modal (A100 GPUs)
Architectures	3D U-Net, Attention U-Net, V-Net, Double U-Net
Augmentation	Albumentations (elastic, geometric, intensity)
Loss	Dice + Focal + Topology-Aware
Strategy	Ensemble + Test-Time Augmentation

Author

Jonathan Muhire — jonathanmuhire.com · GitHub