FedCC uses frozen DINOv2 and LoRA for fetal brain scans with 8.5x less communication

Accurate localization of the corpus callosum in fetal ultrasound images is critical for early detection of neurodevelopmental abnormalities, but remains difficult due to low contrast, speckle noise, and anatomical variability. To address this, researchers from Università Politecnica delle Marche, University of Bologna, and other institutions developed FedCC—a federated learning framework designed for realistic multi-center, resource-constrained clinical environments without requiring data sharing.

FedCC integrates a frozen DINOv2 vision foundation model as a feature extractor with a lightweight YOLO-based detection head. Low-Rank Adaptation (LoRA) modules enable parameter-efficient fine-tuning, allowing only a small subset of parameters to be optimized and exchanged among clients. This reduces trainable parameters to just 2.9 million compared to 24.4 million for full fine-tuning—an 8.5× reduction in communication overhead. Evaluated on 10,970 ultrasound frames from 58 pregnant women across three clinical sites using heterogeneous imaging devices, FedCC achieved an average mAP@50 of 0.857 and an F1-score of 0.803 under the FedAvg strategy, outperforming both full fine-tuning and encoder-freezing baselines. The work represents a promising step toward scalable, privacy-preserving AI systems for fetal neurosonography that can operate in low-resource settings.