ENS × Raidium Data Challenge — 5th Place Solution

Semantic segmentation of 54 anatomical structures from 2D CT scans. Full methodology in Rapport_ENSxRaidium_DataChallenge_Bryan_Ly.pdf.

Results

Model	Dice Score
Custom U-Net + Anchor-Based Cascade	0.5569

Setup

Option A — uv (Recommended)

uv sync
source .venv/bin/activate

Option B — pip

python3.11 -m venv .venv
source .venv/bin/activate
pip install -e .

Data Preparation

Place the raw dataset as follows:

data/raw/
├── train-images/
├── annotated_labels.json
└── label_Hnl61pT.csv

python scripts/data_preprocessing/prepare_data.py
python src/ens_data_challenge/data_processing/make_splits.py

Training

# Phase 1 - Global U-Net
python scripts/train.py

# Phase 2 - Anchor-Based Cascade
python scripts/train_cascade.py

Inference

# Phase 1 - Global U-Net Inference
python scripts/run_inference_no_thresh.py

# Phase 2 - Anchor-Based Cascade Inference
python scripts/run_inference_cascade_final.py