Dante: An Open Source Model Pre-Training and Fine-Tuning Tool for the Dafne Federated Framework for Medical Image Segmentation

Adapting deep learning segmentation models to new clinical settings remains a major hurdle in medical image analysis, especially when annotated data is scarce. Parameter-efficient fine-tuning (PEFT) methods offer a solution by updating only a subset of model weights, preserving learned features while reducing overfitting. A new paper introduces Dante (DAfNe TrainEr), an open-source training backend for the Dafne federated segmentation ecosystem. Dante supports training from scratch with automatic architecture configuration, configurable layer freezing schedules, and a novel extension of Low-Rank Adaptation (LoRA) to N-dimensional convolutional layers via channel-wise factorization. The tool is designed to work seamlessly within federated learning environments, enabling collaborative model adaptation across institutions without centralizing sensitive patient data.

The researchers validated Dante using realistic cross-domain MRI transfer scenarios, including abdominal organ and brain white matter lesion segmentation, under both full-data and few-shot conditions. Two PEFT strategies were tested: Gradual Unfreezing (GU) and LoRA. GU reduced the number of epochs needed to reach 85% of peak performance by up to 63.6% compared to training from scratch, significantly accelerating deployment timelines. LoRA achieved Dice Similarity Coefficients as high as 0.957 in data-rich scenarios, outperforming baseline models across all tested domains. Gains were amplified when using richer pretraining datasets. These results demonstrate Dante's effectiveness as a domain-agnostic fine-tuning module that can be deployed in real clinical settings where annotated data is limited.