Dafne

Deep Anatomical Federated Network is a program for the segmentation of medical images. It relies on a server to provide deep learning models to aid the segmentation, and incremental learning is used to improve the performance. See https://www.dafne.network/ for documentation and user information.

Windows binary installation

Please install the Visual Studio Redistributable Package under windows: https://aka.ms/vs/16/release/vc_redist.x64.exe Then, run the provided installer

Mac binary installation

Install the Dafne App from the downloaded .dmg file as usual. Make sure to download the archive appropriate for your architecture (x86 or arm).

Linux binary installation

The Linux distribution is a self-contained executable file. Simply download it, make it executable, and run it.

pip installation

Dafne can also be installed with pip pip install dafne

Citing

If you are writing a scientific paper, and you used Dafne for your data evaluation, please cite the following paper:

Santini F, Wasserthal J, Agosti A, et al. Deep Anatomical Federated Network (Dafne): an open client/server framework for the continuous collaborative improvement of deep-learning-based medical image segmentation. 2023 doi: 10.48550/arXiv.2302.06352.

Notes for developers

dafne

Run: python dafne.py <path_to_dicom_img>

Notes for the DL models

Apply functions

The input of the apply function is:

dict({
    'image': np.array (2D image)
    'resolution': sequence with two elements (image resolution in mm)
    'split_laterality': True/False (indicates whether the ROIs should be split in L/R if applicable)
    'classification': str - The classification tag of the image (optional, to identify model variants)
})

The output of the classifier is a string. The output of the segmenters is:

dict({
    roi_name_1: np.array (2D mask),
    roi_name_2: ...
})

Incremental learn functions

The input of the incremental learn functions are:

training data: dict({
    'resolution': sequence (see above)
    'classification': str (see above)
    'image_list': list([
        - np.array (2D image)
        - np.array (2D image)
        - ...
    ])
})

training outputs: list([
    - dict({
        roi_name_1: np.array (2D mask)
        roi_name_2: ...
    })
    - dict...

Every entry in the training outputs list corresponds to an entry in the image_list inside the training data. So len(training_data['image_list']) == len(training_outputs).

Acknowledgments

Input/Output is based on DOSMA - GPLv3 license

This software includes the Segment Anything Model (SAM) - Apache 2.0 license

Other packages required for this project are listed in requirements.txt