Paper | Project page

Maximilian Weiherer$^{1,2}$, Antonia von Riedheim$^3$, Vanessa Brébant$^3$, Bernhard Egger$^1$, Christoph Palm$^2$
$^1$ Friedrich-Alexander-Universität Erlangen-Nürnberg
$^2$ OTH Regensburg
$^3$ University Hospital Regensburg

Official implementation of the paper "iRBSM: A Deep Implicit 3D Breast Shape Model", BVM'25.

This repository contains code for the implicit Regensburg Breast Shape Model (iRBSM). Along with the inference code (sampling from our model and reconstructing point clouds), we also provide the code that has been used to train our model. Unfortunately, however, we can't make our training data public.

Abstract: We present the first deep implicit 3D shape model of the female breast, building upon and improving the recently proposed Regensburg Breast Shape Model (RBSM). Compared to its PCA-based predecessor, our model employs implicit neural representations; hence, it can be trained on raw 3D breast scans and eliminates the need for computationally demanding non-rigid registration -- a task that is particularly difficult for feature-less breast shapes. The resulting model, dubbed iRBSM, captures detailed surface geometry including fine structures such as nipples and belly buttons, is highly expressive, and outperforms the RBSM on different surface reconstruction tasks. Finally, leveraging the iRBSM, we present a prototype application to 3D reconstruct breast shapes from just a single image.

We're using Python 3.9, PyTorch 2.0.1, and CUDA 11.7. To install all dependencies within a conda environment, simply run:

conda env create -f environment.yaml
conda activate irbsm

This may take a while.

You can download the iRBSM here. After downloading, make sure to place the .pth file in the ./checkpoints folder.

To produce random samples from the iRBSM, use

python sample.py <number-of-samples>

This generates <number-of-samples> random breast shapes which are saved as .ply files.

Optional arguments:

• --output_dir: The output directory where to save samples. Default: ./.
• --device: The device the model should run on. Default: cuda.
• --voxel_resolution: The resolution of the voxel grid on which the implicit model is evaluated. Default: 256.
• --chunk_size: The size of the chunks the voxel grid should be split into. Default: 100_000. If you have a small GPU with only little VRAM, lower this number.

To reconstruct a point cloud using the iRBSM, type

python reconstruct.py <path-to-point-cloud>

Optional arguments:

• --landmarks: Path to a file containing landmarks. Please see details below.
• --output_dir: The output directory where to save the reconstruction. Default: ./.
• --latent_weight: The regularization weight that penalizes the L2 norm of the latent code. Default: 0.01. If you have noisy inputs, increase this number.
• --device: The device the model should run on. Default: cuda.
• --voxel_resolution: The resolution of the voxel grid on which the implicit model is evaluated. Default: 256.
• --chunk_size: The size of the chunks the voxel grid should be split into. Default: 100_000. If you have a small GPU with only little VRAM, lower this number.

Whenever the orientation of the given point cloud differs significantly from the model's orientation, you should first roughly align both coordinate systems (this is necessary because our model is not invariant to rigid transformations). The easiest way to achieve this is by providing certain landmark positions. These points can then be used to rigidly align the given point cloud to the model. Please provide the following landmarks in exactly this order:

1. Sternal notch
2. Belly button
3. Left nipple (from the patient's perspective; so it's actually right from your perspective!)
4. Right nipple (from the patient's perspective; so it's actually left from your perspective!)

We recommend using MeshLab's PickPoints (PP) tool, which allows you to export selected point positions as XML file with .pp extension. You can directly pass this file to reconstruct.py. Alternatively, you can use your favorite point picker tool and pass points as comma-separated .csv file. Lastly, we also provide a simple application to interactively select points, just run

python scripts/pick_landmarks.py <path-to-point-cloud>

Please also see the README file in ./scripts.

To train our model on your own watertight 3D breast scans, you first need to bring your data into the file format we're using (we're expecting training data to be stored in .hdf5 files). The following script does that for you; it first scales raw meshes into the unit cube, and then optionally discards inner structures. Finally, it produces a ready-to-use .hdf5 file that you can later plug into our training pipeline. Simply type

python scripts/preprocess_dataset.py <path-to-your-scans>

Optional arguments:

• --output: The name of the output .hdf5 file. Default: ./dataset.hdf5.
• --padding: The padding to add to the unit cube. Default: 0.1.

After preprocessing, make sure to place the resulting .hdf5 file(s) in the ./data folder.

To train the model, type

python train.py configs/deep_sdf_256.yaml

For training, you'd also need a wandb account. To log in to your account, simply type wandb login and follow the instructions.

If you use the iRBSM, please cite

@inproceedings{weiherer2025irbsm,
    title={iRBSM: A Deep Implicit 3D Breast Shape Model},
    author={Weiherer, Maximilian and von Riedheim, Antonia and Brébant, Vanessa and Egger, Bernhard and Palm, Christoph},
    booktitle={Proceedings of the German Conference on Medical Image Computing},
    pages={38–-43},
    year={2025}
}

and

@article{weiherer2023rbsm,
  title={Learning the shape of female breasts: an open-access 3D statistical shape model of the female breast built from 110 breast scans},
  author={Weiherer, Maximilian and Eigenberger, Andreas and Egger, Bernhard and Brébant, Vanessa and Prantl, Lukas and Palm, Christoph},
  journal={The Visual Computer},
  volume={39},
  pages={1597–-1616},
  year={2023}
}

Also, in case you have any questions, feel free to contact Maximilian Weiherer, Bernhard Egger, or Christoph Palm.