🤖 machine learning

FedCC: A Low-Resource Federated Adaptation of Foundation Models for Robust Corpus Callosum localization in Fetal Ultrasound Images

FedCC is a low-resource federated learning framework that combines a frozen DINOv2 backbone with LoRA and a YOLO-based detection head to achieve robust, privacy-preserving localization of the corpus callosum in fetal ultrasound images while significantly reducing computational and communication costs compared to full fine-tuning.

Original authors: Alessandro Di Matteo, Sara Moccia, Giuseppe Rizzo, Gianpaolo Grisolia, Ricciarda Raffaelli, Lorenzo Vasciaveo, Francesco D'Antonio, Maria Chiara Fiorentino

Published 2026-07-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alessandro Di Matteo, Sara Moccia, Giuseppe Rizzo, Gianpaolo Grisolia, Ricciarda Raffaelli, Lorenzo Vasciaveo, Francesco D'Antonio, Maria Chiara Fiorentino

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a super-smart robot how to spot a tiny, hidden treasure in a foggy, grainy photograph. This isn't just any photo; it's an ultrasound of a baby's brain, where the "treasure" is a delicate, curved structure called the corpus callosum. This structure is the bridge connecting the two halves of the brain, and finding it early is crucial for spotting developmental issues. But here's the catch: ultrasound images are notoriously messy. They are full of static (like an old TV), low contrast, and the baby moves around, making the picture look different every time.

Now, imagine you want to teach this robot using photos from three different hospitals, but the hospitals can't share their actual patient photos because of strict privacy rules. It's like trying to solve a giant jigsaw puzzle where everyone has their own pieces, but no one is allowed to show their pieces to anyone else. Usually, to get a robot this smart, you'd need to send all those photos to one giant computer, which breaks the privacy rule. Or, you'd have to send the robot's entire "brain" (its complex mathematical model) back and forth between hospitals, which is like mailing a heavy, fragile library book every time you want to add a single new fact. This paper asks: Is there a way to teach the robot to find that tiny brain bridge without moving the heavy books or the private photos?

The researchers behind this study, who call their method FedCC, say yes. They built a clever system that combines three big ideas. First, they use a "foundation model" called DINOv2. Think of this as a robot that has already read millions of books and seen billions of pictures of the world; it already knows what shapes, edges, and textures look like, so it doesn't need to start from scratch. Second, they use a technique called LoRA (Low-Rank Adaptation). If the foundation model is a giant library, LoRA is like giving the robot a tiny, lightweight notebook. Instead of rewriting the whole library to learn about fetal brains, the robot just writes a few new notes in its notebook. This notebook is so small it's easy to carry. Third, they use Federated Learning. This is the magic trick where the robot learns locally at each hospital, writes its notes in its tiny notebook, and only sends those small notebooks to a central server to be combined. The heavy library stays locked in each hospital, and the private photos never leave the building.

The team tested this idea on a real-world challenge: finding the corpus callosum in ultrasound images from three different Italian hospitals. These hospitals used different machines, had different image settings, and different types of patients, creating a "messy" mix of data that usually confuses robots. They simulated a scenario where the hospitals couldn't share their raw data.

What they found is quite promising. By using their "tiny notebook" (LoRA) approach with the pre-trained "library" (DINOv2), the system learned to spot the brain bridge with an average accuracy (called mAP@50) of 0.857 and a balance score (F1-score) of 0.803. This was better than trying to rewrite the whole library (full fine-tuning) or just freezing the library and only training the very end of the robot.

The most exciting part is the efficiency. Because they only sent the tiny notebooks instead of the whole library, they reduced the amount of data that needed to be sent between hospitals by 8.5 times. In numbers, they only had to train and share 2.9 million parameters (the robot's "brain weights"), whereas the old way would have required moving 24.4 million parameters. This suggests that you can build a powerful, privacy-safe medical AI that works across different hospitals without needing a supercomputer or breaking privacy laws.

However, the authors are careful to note that this is a simulation based on data from 58 pregnant women across three centers, and all the cases were normal pregnancies without any brain abnormalities. The system is great at finding the bridge, but it hasn't been tested yet on babies with actual defects. Also, the system currently draws a box around the structure rather than tracing its exact shape, so it can't yet measure the bridge's length or thickness.

In short, the paper suggests that by using a smart, pre-trained robot that only writes tiny notes to learn new tasks, we can teach medical AI to work across different hospitals while keeping patient data safe and saving massive amounts of computing power. It's a step toward a future where hospitals can collaborate on AI without ever having to share their most sensitive secrets.

Technical Summary: FedCC – A Low-Resource Federated Adaptation of Foundation Models for Robust Corpus Callosum Localization

Problem Statement
Accurate localization of the corpus callosum (CC) in fetal ultrasound (US) images is critical for the early detection of neurodevelopmental abnormalities. However, this task faces significant challenges due to the intrinsic limitations of US imaging, including low contrast, speckle noise, and substantial anatomical variability. Furthermore, the development of robust automated systems is hindered by inter-site variability caused by heterogeneous acquisition protocols and imaging devices. Traditional deep learning approaches often require large, centralized datasets, which conflict with privacy regulations and ethical constraints preventing data sharing between institutions. Additionally, adapting large Foundation Models (FMs) to specific medical tasks typically involves full fine-tuning, which imposes prohibitive computational and communication costs, particularly in resource-constrained clinical environments.

Methodology
The authors propose FedCC, a federated learning (FL) framework designed for multi-center, resource-constrained settings that operates without exchanging raw patient data. The framework integrates three core components:

Frozen Foundation Backbone: The system utilizes DINOv2, a Vision Transformer (ViT) pretrained via self-supervised learning on large-scale natural images. Crucially, the backbone weights ( $\omega$ ) remain frozen throughout the training process and are never transmitted between clients and the server. This preserves the rich, generalizable representations learned during pretraining.
Parameter-Efficient Adaptation (LoRA): To adapt the frozen backbone to the specific task of CC localization, the authors incorporate Low-Rank Adaptation (LoRA) modules. These modules are inserted into the query and value projection matrices of the transformer attention blocks. Instead of updating the full weight matrix $W$ , LoRA constrains updates to a low-rank residual ( $\Delta W = BA$ ), where $r \ll \min(d, k)$ . This drastically reduces the number of trainable parameters.
Lightweight Detection Head: The latent representations from the backbone are processed by a compact, single-scale detection head derived from the YOLOv8 architecture. This branch includes Spatial Pyramid Pooling–Fast (SPPF) and C3f blocks to refine features and predict bounding boxes and confidence scores directly, avoiding the complexity of multi-stage region proposal networks.

Federated Training Strategy
The framework employs a standard FL protocol where $N$ clinical centers (clients) collaboratively train a shared model.

Local Training: Each client trains only the LoRA adapters ( $\alpha_i$ ) and the detection head parameters ( $\phi_i$ ) on local data for $E$ epochs. The backbone remains fixed locally.
Communication: Only the lightweight trainable parameters $\theta_i = \{\alpha_i, \phi_i\}$ are transmitted to the central server.
Aggregation: The server aggregates these parameters using standard strategies (FedAvg or FedProx) and redistributes the updated global parameters to clients for the next round.
Privacy & Efficiency: Since raw images, annotations, and backbone weights are never exchanged, the approach ensures data privacy. The reduction in transmitted parameters significantly lowers communication overhead.

Key Contributions

FedCC Framework: A novel FL framework combining a frozen DINOv2 backbone with LoRA adapters and a YOLO-based detection head for CC localization. This design enables the practical adaptation of large FMs in distributed clinical settings while preserving privacy.
Systematic Evaluation of Robustness: The study evaluates the framework under realistic multi-center and multi-device scenarios, characterized by significant domain shifts in scanner types, resolution, and image quality. Comparisons against full fine-tuning, encoder freezing, and various aggregation methods demonstrate that the proposed LoRA-based federated adaptation improves cross-client generalization.
Novel Dataset: The authors introduce a new fetal US dataset for CC detection comprising 10,970 frames from 58 pregnant women across three Italian clinical centers (Chieti, Foggia, Verona), utilizing heterogeneous imaging devices (GE Voluson E10, Samsung HERA W9/W10/Z20).

Results
The proposed framework was evaluated on the multi-center dataset using metrics including mAP@50, Precision, Recall, and F1-score.

Performance: The combination of DINOv2base (natural image pretraining) with the proposed LoRA strategy under FedAvg achieved the best results, with an average mAP@50 of 0.857 and an F1-score of 0.803. This outperformed both full fine-tuning and encoder-freezing baselines.
Generalization: The federated approach demonstrated superior generalization across heterogeneous domains, particularly on the most challenging client (Client 3), where centralized training on pooled data degraded significantly (dropping to 0.422 mAP@50 on that client).
Efficiency: The proposed method reduced the number of trainable parameters to 2.9M (compared to 24.4M for full fine-tuning), resulting in an approximately 8.5× reduction in communication costs.
Comparisons: The proposed approach outperformed other foundation models (SAM, UltraSAM, UltraFedFM) and CNN-based YOLO baselines (YOLO26 variants) in the federated setting. Notably, natural-image pretraining (DINOv2base) proved more effective than fetal-specific pretraining (DINOv2fetal) in this context.

Significance and Claims
The paper claims that FedCC represents a promising step toward scalable, privacy-preserving, and clinically deployable AI systems for fetal neurosonography. By integrating Federated Learning, Foundation Models, and Parameter-Efficient Fine-Tuning (PEFT), the framework addresses the triad of challenges: privacy preservation, inter-scanner domain shift, and communication efficiency.

The authors modestly note that while the results are strong, further validation on larger and more diverse cohorts is required before clinical translation. They acknowledge limitations, including the exclusion of CC malformations from the dataset (focusing only on localization in normal pregnancies) and the simulation of the FL environment under controlled conditions (synchronous updates). Future work is proposed to extend the framework to automated CC segmentation for biometric quantification and to detect other anatomical subregions.

Technical Summary: FedCC – A Low-Resource Federated Adaptation of Foundation Models for Robust Corpus Callosum Localization

More like this