FedCC Localizes Fetal Brain Structures with Low-Resource Federated Learning

Accurate identification of the corpus callosum (CC) in fetal ultrasound images is critical for detecting early neurodevelopmental issues. However, challenges like low contrast, speckle noise, and anatomical variability make this a difficult task. This research introduces FedCC, a federated learning (FL) framework tailored for realistic clinical environments with limited resources and strict data privacy requirements. FedCC employs a robust architecture that integrates a pre-trained DINOv2 backbone, which remains frozen, with a lightweight YOLO-based detection head. To ensure efficient adaptation without sharing raw patient data, the framework incorporates Low-Rank Adaptation (LoRA) modules. LoRA allows only a small fraction of parameters to be optimized and exchanged between participating clinical sites, drastically cutting down on computational and communication costs. Evaluated on a multi-center dataset of nearly 11,000 ultrasound frames from 58 pregnant women across three sites, FedCC demonstrated strong performance. The DINOv2 and LoRA combination, under a FedAvg strategy, achieved an mAP@50 of 0.857 and an F1-score of 0.803. This approach significantly outperformed full fine-tuning and encoder-freezing baselines while reducing trainable parameters by approximately 8.5 times, making it highly suitable for privacy-preserving and scalable clinical AI deployments.