This paper introduces DAfNe TrainEr (Dante), an open-source training and fine-tuning module designed to help local sites adapt pretrained medical image segmentation models to their own MRI data. Dante plugs into the Dafne federated segmentation system, which lets hospitals share model improvements without sharing raw patient images. The goal is to let a site start from a model that already “knows” the anatomy and update only the parts that need to change for local scanners and protocols.

The authors built Dante to do two main things. First, it can train common segmentation networks from scratch using three architectures: 2D U-Net, 3D U-Net, and DynUNet. DynUNet follows the self-configuration idea of nnU-Net, meaning it automatically sets network and training parameters from the dataset’s voxel spacing and image size. Second, Dante offers parameter-efficient fine-tuning methods so a model can adapt with few local annotations. These include Layers Freezing with a Gradual Unfreezing schedule (which keeps low-level features fixed and progressively allows deeper layers to change) and Low-Rank Adaptation (LoRA), a method that represents weight updates with small, low-rank matrices and here is extended to N-dimensional convolutional layers using channel-wise factorization.

To test Dante, the researchers ran transfer experiments on public MRI tasks: abdominal organ segmentation and brain white matter lesion segmentation. They compared training from scratch to Gradual Unfreezing and LoRA under both full-data and few-shot conditions. The paper reports that Gradual Unfreezing cut the number of training epochs needed to reach 85% of peak performance by up to 63.6% compared to training from scratch. LoRA reached Dice Similarity Coefficients (a common measure of segmentation overlap where 1.0 is perfect) as high as 0.957 in data-rich settings. Both strategies outperformed their baseline in every tested domain, and they worked even better when the starting model was pretrained on larger, richer datasets.