MIC-DKFZ / RetinaUNet-MissingPiece-MultiTalent

RetinaUNet-MissingPiece-MultiTalent

_{Copyright German Cancer Research Center (DKFZ) and contributors. Please make sure that your usage of these models is in compliance with their license.}

RetinaUNet-MissingPiece-MultiTalent is a Retina U-Net backbone for 3D medical object detection , pre-trained with MultiTalent supervised multi-dataset segmentation on ~87 public segmentation datasets of mixed modalities. It is one of the checkpoints released with The Missing Piece: A Case for Pre-training in 3D Medical Object Detection (MICCAI 2025).

Model family

All checkpoints of the collection :

The nnFoundation models nnFoundationCNN (ResEnc-L) and nnFoundationViT (Primus) can be finetuned for detection with nnDetection in the same way -- see the finetuning docs .

Using this checkpoint

Fine-tune it for detection with nnDetection -- Finetuning pretrained backbones :

hf download MIC-DKFZ/RetinaUNet-MissingPiece-MultiTalent checkpoint_final.pth --local-dir ./checkpoints/RetinaUNet-MissingPiece-MultiTalent

nndet_train Task<XXX>_YourDataset retinaunet_focal_v002_for_ConvBackboneMultiTalent 0 \
    -o exp.tag=_MultiTalentRetinaUNet \
    +transfer_learning_ckpt=./checkpoints/RetinaUNet-MissingPiece-MultiTalent/checkpoint_final.pth \
    --transfer_learning --load_adapt_plan

--load_adapt_plan rebuilds the backbone to match this checkpoint before loading the weights. This checkpoint contains the Retina U-Net encoder and FPN decoder, without the detection heads.

Repository contents

Expected input

3D volumes, preprocessed with nnDetection's standard pipeline ( nndet_prep ), using the dataset's default planned spacing . Downstream patch size: 128 × 128 × 128 .

For reference, pre-training used Z-score normalised volumes resampled to 1 × 1 × 1 mm, with a patch size of 128 × 128 × 128.

The checkpoint contains 87 per-dataset input stems (one per pre-training dataset/modality). The adaptation plan selects the CT stem ( stem.351_0 ) by default; to use a different one, pass -o model_cfg.stem_override=stem.<id> (see §3.4 of the finetuning docs ).

Checkpoint format

checkpoint_final.pth is a torch.save dictionary that loads safely with weights_only=True :

Citation

If you use these weights, please cite The Missing Piece (MICCAI 2025):

@inproceedings{eckstein2025missing,
  title     = {The Missing Piece: A Case for Pre-training in 3D Medical Object Detection},
  author    = {Eckstein, Katharina and Ulrich, Constantin and Baumgartner, Michael and K{\"a}chele, Jessica and Bounias, Dimitrios and Wald, Tassilo and Floca, Ralf and Maier-Hein, Klaus H.},
  booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
  series    = {Lecture Notes in Computer Science},
  volume    = {15963},
  pages     = {615--626},
  year      = {2025},
  publisher = {Springer Nature Switzerland},
  doi       = {10.1007/978-3-032-04965-0_58}
}

Please also cite the architecture, pre-training method and pre-training data behind this checkpoint: