Dafne (Deep Anatomical Federated Network) is a collaborative platform to annotate MRI images and train machine learning models without your data ever leaving your machine.


Francesco Santini1 (Project coordination and UI), Jakob Wasserthal1 (Client/Server architecture), Abramo Agosti1 (Deep Learning architecture), Anna Pichiecchio2 (Medical advisor)

Estimated cost




Dafne - Deep Anatomical Federated Network

Dafne is a program for the segmentation of medical images, specifically MR images, that includes advanced deep learning models for an automatic segmentation. The user has the option of refining the automated results, and the software will learn from the improvements and modify its internal models accordingly. In order to continuously improve the performance, the deep learning modules are stored in a central server location.

The main feature of Dafne is the possibility of doing automatic segmentation using deep learning models.

Dafne uses incremental learning and federated learning to continuously adapt the models to the need of our users. This means that when you perform a segmentation, initially it will not be perfect. You will then have the chance to refine it. When you are satisfied with your dataset, you will export your ROI masks. During this export procedure, the software automatically learns from your refined segmentation and sends the updated model back to our servers. We will automatically integrate your updated model with the models of our other users, so you will always receive the most accurate predictor.

It is important to say that your data never leave your computer (unless you explicitly want to), so you don’t have to worry about privacy and data safety.

Currently, we offer two deep learning models:

  1. The thigh model, which identifies 12 muscles of the thigh.
  2. The leg model, which identifies 6 muscles of the leg.

Both models are pretrained on axial proton-density-weighted in-phase gradient echo images. Best results will be achieved with similar protocols; however, the model will adapt in the future depending on the user base.

Video tutorials:

  1. Introduction
  2. Usage
  3. Development


1University of Basel, Basel, Switzerland

2Istituto di Ricovero e Cura a Carattere Scientifico Mondino Foundation, Pavia, Italy