---
license: apache-2.0
tags:
  - medical-imaging
  - segmentation
  - nnunet
  - ct
datasets:
  - FLARE-MedFM/FLARE-Task1-Pancancer
---

# PanCancerSeg Specialized Weights

Cancer-specific nnUNet v2 segmentation models trained on CT images from the [CVPR 2026 FLARE Task 1: Pan-cancer Segmentation](https://www.codabench.org/competitions/7149/) dataset.

## Models

| Folder | Cancer type |
|--------|-------------|
| Dataset102_Kidney | Kidney cancer |
| Dataset103_Liver | Liver cancer |
| Dataset104_Pancreas | Pancreatic cancer |
| Dataset105_Lung | Lung cancer |

All models use `nnUNetTrainerWandb2000`, `nnUNetResEncUNetMPlans`, `3d_fullres`, fold 0, trained for 2000 epochs. Each folder contains `checkpoint_best.pth` (best validation checkpoint).

## Usage

Download the weights and point `--model_dir` at the root directory:

```bash
# Clone this repo
git lfs install
git clone https://huggingface.co/KS987/PanCancerSeg-Specialized-weights

# Run inference
python predict.py \
  --input /path/to/case.nii.gz \
  --cancer_type kidney_cancer \
  --model_dir ./PanCancerSeg-Specialized-weights \
  --device cuda
```

See [PanCancerSeg-Inference](https://github.com/Kappapapa123/PanCancerSeg-Inference) for the full inference pipeline.

## File Structure

```
Dataset10X_*/
└── nnUNetTrainerWandb2000__nnUNetResEncUNetMPlans__3d_fullres/
    ├── dataset.json
    ├── dataset_fingerprint.json
    ├── plans.json
    └── fold_0/
        └── checkpoint_best.pth
```