total-feet-seg

This repository contains nnU-Net v2 models that segment 56 bones of both feet and ankles in CT scans: the distal tibia and fibula, the tarsals, the metatarsals and the phalanges. Left and right bones get separate labels.

Models

Model	Plans (-p)	Epochs	Pseudo-Dice	Size
ResEnc L U-Net (recommended)	nnUNetResEncUNetLPlans	681	0.958	821 MB
Plain U-Net (smaller, faster)	nnUNetPlans	179	0.940	250 MB

Pseudo-Dice by bone group:

Bone group	ResEnc L U-Net	Plain U-Net
Tibia and fibula	0.978	0.974
Tarsals	0.975	0.958
Metatarsals	0.966	0.959
Phalanges	0.944	0.919

The pseudo-Dice is nnU-Net's validation score during training. Both models were trained on all 30 cases (fold all), so the score is measured on training data and does not estimate accuracy on new scans.

Both models were trained with these settings:

• Trainer: nnUNetTrainerNoMirroring. Mirroring is turned off, so left and right are never swapped.
• Configuration: 3d_fullres
• Fold: all
• Checkpoint: checkpoint_best.pth

Training data

The models were trained on 30 CT scans of the feet. The images come from VSDFullBodyBoneReconstruction. The labels were created from the bone meshes of VSDFullBodyBoneModels, with the fused foot bones split into individual bones.

The training data is published as BoneHub/vsd-feet-seg.

Labels

	Left	Right
Tibia	1	2
Fibula	3	4
Talus	5	6
Calcaneus	7	8
Navicular	9	10
Cuboid	11	12
Medial / Intermediate / Lateral cuneiform	13 / 15 / 17	14 / 16 / 18
Metatarsals 1–5	19, 21, 23, 25, 27	20, 22, 24, 26, 28
Phalanges of toe 1 (proximal, distal)	29, 31	30, 32
Phalanges of toes 2–5 (proximal, middle, distal)	33–55, odd	34–56, even

The full list of label names is in models/*/dataset.json. In a name like Phalange_Foot_<toe>_<phalanx>_<side>:

• Toe: 1 is the big toe.
• Phalanx: counted from proximal to distal.

Prediction

1. Install

pip install nnunetv2 huggingface_hub

Install PyTorch for your CUDA version first.

2. Download the models and rename the folder

nnU-Net expects trained models in $nnUNet_results/DatasetXXX_Name/. The models/ folder in this repo therefore has to be renamed after download, for example to Dataset001_total-feet-seg.

export nnUNet_results="$PWD/nnUNet_results"      # Windows: $env:nnUNet_results = "$PWD\nnUNet_results"

hf download BoneHub/total-feet-seg --include "models/*" --local-dir ./total-feet-seg
mkdir -p "$nnUNet_results"
mv ./total-feet-seg/models "$nnUNet_results/Dataset001_total-feet-seg"

If you already have another Dataset001_*, choose a free ID (e.g. Dataset901_total-feet-seg) and use that number with -d below.

3. Prepare the input

Put your CT scans (in Hounsfield units, .nii.gz) in one folder. Each file name must end in _0000:

data/infer/
├── subject01_0000.nii.gz
└── subject02_0000.nii.gz

The image orientation must be correct, because the model predicts left and right labels.

4. Predict

ResEnc L U-Net:

nnUNetv2_predict -i ./data/infer -o ./data/infer_out -d 1 -tr nnUNetTrainerNoMirroring -p nnUNetResEncUNetLPlans -c 3d_fullres -f all -chk checkpoint_best.pth

Plain U-Net:

nnUNetv2_predict -i ./data/infer -o ./data/infer_out -d 1 -tr nnUNetTrainerNoMirroring -p nnUNetPlans -c 3d_fullres -f all -chk checkpoint_best.pth

Each output in ./data/infer_out/ is a label map with the same geometry as its input.

Retraining

1. Download the dataset and rename the folder

The dataset folder feet_ct/ has to be renamed to an nnU-Net dataset name inside $nnUNet_raw, for example Dataset002_vsd-feet-seg.

Use a different ID than the downloaded model (Dataset001_total-feet-seg). If both use the same ID, training writes into the model's folder and overwrites the released weights.

export nnUNet_raw="$PWD/nnUNet_raw"
export nnUNet_preprocessed="$PWD/nnUNet_preprocessed"
export nnUNet_results="$PWD/nnUNet_results"

hf download BoneHub/vsd-feet-seg --repo-type dataset --include "feet_ct/*" --local-dir ./vsd-feet-seg
mkdir -p "$nnUNet_raw"
mv ./vsd-feet-seg/feet_ct "$nnUNet_raw/Dataset002_vsd-feet-seg"

2. Preprocess and train

ResEnc L U-Net:

nnUNetv2_plan_and_preprocess -d 2 -pl nnUNetPlannerResEncL -c 3d_fullres --verify_dataset_integrity
nnUNetv2_train 2 3d_fullres all -tr nnUNetTrainerNoMirroring -p nnUNetResEncUNetLPlans

Plain U-Net:

nnUNetv2_plan_and_preprocess -d 2 -c 3d_fullres --verify_dataset_integrity
nnUNetv2_train 2 3d_fullres all -tr nnUNetTrainerNoMirroring -p nnUNetPlans

• Always use nnUNetTrainerNoMirroring. The default trainer mirrors images during training, which mixes up left and right.
• To estimate accuracy on unseen data, train folds 0–4 instead of all.
• Predicting with the retrained model. Use -d 2 and leave out -chk. A finished training run saves checkpoint_final.pth, which nnU-Net uses by default.

Fine-tuning on your own data

This section starts from the released ResEnc L weights.

1. Prepare your data. Create your own dataset in $nnUNet_raw, e.g. Dataset004_my-feet, using the nnU-Net format (imagesTr/, labelsTr/, dataset.json). Use the same labels as above if possible.

2. Reuse the released plans and preprocess. Copy the model's plans to $nnUNet_preprocessed, transfer them to your dataset, and preprocess:

   M="$nnUNet_results/Dataset001_total-feet-seg/nnUNetTrainerNoMirroring__nnUNetResEncUNetLPlans__3d_fullres"
   mkdir -p "$nnUNet_preprocessed/Dataset001_total-feet-seg"
   cp "$M/plans.json" "$nnUNet_preprocessed/Dataset001_total-feet-seg/nnUNetResEncUNetLPlans.json"
   
   nnUNetv2_extract_fingerprint -d 4
   nnUNetv2_move_plans_between_datasets -s 1 -t 4 -sp nnUNetResEncUNetLPlans -tp nnUNetResEncUNetLPlans
   nnUNetv2_preprocess -d 4 -plans_name nnUNetResEncUNetLPlans -c 3d_fullres
   

3. Train from the released weights.

   nnUNetv2_train 4 3d_fullres all -tr nnUNetTrainerNoMirroring -p nnUNetResEncUNetLPlans \
       -pretrained_weights "$M/fold_all/checkpoint_best.pth"
   

See nnU-Net's fine-tuning guide for details.

Limitations

• The models were trained on only 30 subjects and not tested on an independent dataset.
• Results may be worse on other scanners, fields of view, implants or pathologies.
• For research use only. The models are not for clinical use.

License

CC BY-NC-SA 4.0

Citation

If you use these models, please cite nnU-Net and the used datasets:

@article{isensee2021nnunet,
  title   = {nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation},
  author  = {Isensee, Fabian and Jaeger, Paul F. and Kohl, Simon A. A. and Petersen, Jens and Maier-Hein, Klaus H.},
  journal = {Nature Methods},
  volume  = {18},
  number  = {2},
  pages   = {203--211},
  year    = {2021}
}

@misc{seyedhamidrezaalaviVsdfeetseg,
  title = {Vsd-Feet-Seg},
  author = {Seyed Hamidreza Alavi and Malte Asseln},
  year = {2026},
  publisher = {Hugging Face},
  doi = {10.57967/HF/10470},
  url = {https://huggingface.co/datasets/BoneHub/vsd-feet-seg},
  urldate = {2026-09-16}
}

@article{fischerDatabaseSegmentationsSurface2023,
  title = {Database of Segmentations and Surface Models of Bones of the Entire Lower Body Created from Cadaver {{CT}} Scans},
  author = {Fischer, M.C.M.},
  year = 2023,
  journal = {Scientific Data},
  volume = {10},
  number = {1},
  pages = {763},
  publisher = {Nature Publishing Group},
  issn = {2052-4463},
  doi = {10.1038/s41597-023-02669-z},
}