---
license: apache-2.0
tags:
  - medical-imaging
  - segmentation
  - nnunet
  - computed-tomography
  - cancer
  - cvpr2026
  - flare
datasets:
  - FLARE-MedFM/FLARE-Task1-Pancancer
metrics:
  - dice
  - normalized-surface-dice
pipeline_tag: image-segmentation
---

# Vanilla nnUNet v2 Baseline — CVPR 2026 Task 1: Pan-Cancer Segmentation

Binary lesion segmentation from CT using vanilla nnU-Net v2 (ResEncM). Reference baseline for [CVPR 2026: Foundation Models for Pan-cancer Segmentation in CT Images Task 1](https://www.codabench.org/competitions/7149/).

**Code + failure analysis:** [GitHub — Kappapapa123/CVPR2026-PanCancerSeg-baseline](https://github.com/Kappapapa123/CVPR2026-PanCancerSeg-baseline)

---

## Model Details

| Parameter | Value |
|-----------|-------|
| Framework | nnU-Net v2.6.4 |
| Architecture | Residual Encoder U-Net Medium (`nnUNetResEncUNetMPlans`) |
| Configuration | 3d_fullres |
| Task | Binary segmentation (0=background, 1=tumor) |
| Epochs | 2000 |
| Fold | 0 (official competition split via `splits_final.json`) |
| Trainer class | `nnUNetTrainerWandb2000` |
| Dataset ID | 101 (`Dataset101_FLAREPanCancer`) |

## Training Data

[FLARE-MedFM/FLARE-Task1-Pancancer](https://huggingface.co/datasets/FLARE-MedFM/FLARE-Task1-Pancancer) — 10,762 cases total:
- **train_label** (~10,000 cases): Partially labeled (primary lesion only per case)
- **DeepLesion5K-MedSAM2** (~762 cases): Pseudo labels from MedSAM2

Only the primary lesion is annotated per training case. Metastatic lesions may be present but unlabeled.

## Performance

Evaluated on 50 validation-public cases:

| Metric | Mean | Median | Std |
|--------|------|--------|-----|
| Lesion DSC | 0.3278 | 0.2318 | 0.2883 |
| Lesion NSD (2 mm) | 0.1797 | 0.1046 | 0.2029 |

## How to Use

```bash
# 1. Install dependencies
pip install nnunetv2==2.6.4 huggingface-hub

# 2. Clone code repo (needed for the custom trainer)
git clone https://github.com/Kappapapa123/CVPR2026-PanCancerSeg-baseline.git

# 3. Symlink the custom trainer into nnUNet
VARIANTS_DIR=$(python -c "import nnunetv2; print(nnunetv2.__path__[0])")/training/nnUNetTrainer/variants
ln -sf $(realpath CVPR2026-PanCancerSeg-baseline/trainers/nnUNetTrainerWandb2000.py) \
    "$VARIANTS_DIR/nnUNetTrainerWandb2000.py"

# 4. Download model weights
huggingface-cli download KS987/CVPR2026-PanCancerSeg-baseline \
    --local-dir ./model --repo-type model

# 5. Run inference
nnUNetv2_predict \
    -i /path/to/input_images \
    -o /path/to/output_predictions \
    -d 101 -c 3d_fullres -f 0 \
    -tr nnUNetTrainerWandb2000 \
    -p nnUNetResEncUNetMPlans \
    -m ./model/nnUNet/Dataset101_FLAREPanCancer \
    --disable_tta
```

Both `checkpoint_final.pth` and `checkpoint_best.pth` are provided. nnUNet uses `checkpoint_final.pth` by default.

> Inference resampling requires ~64 GB RAM.

## Limitations

- **Partial labels:** Training data only labels the primary lesion per case.
- **False positive fragmentation:** The dominant failure mode is scattered false-positive blobs. See the [failure analysis](https://github.com/Kappapapa123/CVPR2026-PanCancerSeg-baseline#failure-analysis).

## Files

```
nnUNet/
  Dataset101_FLAREPanCancer/
    nnUNetTrainerWandb2000__nnUNetResEncUNetMPlans__3d_fullres/
      fold_0/
        checkpoint_best.pth      # Best validation checkpoint
        checkpoint_final.pth     # Final epoch (2000) checkpoint
      dataset.json
      dataset_fingerprint.json
      plans.json
```