Abstract
Federated Medical Image Segmentation (FMIS) allows collaborative model training across institutions without sharing raw data. However, medical annotations are often noisy and biased due to varying expertise and subjective judgment. FedA3I addresses this by introducing a quality-aware aggregation strategy that handles "heterogeneous annotation noise." By modeling realistic contour shifts and identifying high-quality clients through early learning dynamics, FedA3I prioritizes clean data for semantic layers while leveraging noisy data for low-level feature extraction.
Introduction
Federated learning is a promising paradigm for medical imaging, but it faces a significant challenge: medical datasets are inherently noisy. Unlike standard machine learning assumptions where noise is random (IID), medical annotation noise is typically Non-IID at both the pixel and client levels.
- Pixel Level: Errors are concentrated around object boundaries rather than being randomly distributed.
- Client Level: Different institutions exhibit systematic biases—some consistently over-segment while others under-segment. Standard aggregation methods like FedAvg, which weight clients by data quantity, are vulnerable to "poisoning" from large, poorly annotated datasets.
Methodology
To address these challenges, the authors propose two main innovations: the Contour Evolution Model (CEM) and the FedA3I aggregation framework.
The Contour Evolution Model (CEM)
To simulate realistic medical noise, the authors developed CEM. Instead of simple label-flipping, CEM treats annotations as contours that evolve inward or outward based on systematic preference (bias) and irregular distortions (randomness). This creates a more realistic benchmark for testing noise-robust algorithms.
FedA3I Framework
The FedA3I framework operates in three stages:
- Noise Estimation via Early Learning: Utilizing the "early learning" phenomenon where networks learn clean patterns before memorizing noise, the model uses initial loss values to estimate client data quality.
- Client Grouping: A Gaussian Mixture Model (GMM) clusters clients based on their segmentation tendencies (over- vs. under-segmenting).
- Layer-wise Aggregation: Recognizing that noisy labels impact deep semantic layers more than shallow edge-detection layers, FedA3I applies dynamic weighting. Shallow layers use standard quantity-based weights, while deeper layers are weighted toward clients with higher-quality annotations.
Results and Discussion
The method was evaluated on the ISIC 2017 skin lesion (SKIN) and breast ultrasound (BREAST) datasets. FedA3I consistently outperformed state-of-the-art noise-robust methods. On the SKIN dataset with heavy noise, FedA3I achieved a Dice coefficient of 72.69%, significantly higher than FedNoRo (66.35%).
Ablation studies confirmed that "Inter-group weighting"—balancing over- and under-segmenting groups—is vital to prevent the global model from drifting toward a majority bias. Furthermore, FedA3I adds almost zero computational overhead to the local training process, making it highly efficient for real-world deployment.
Conclusion
FedA3I provides a robust solution for federated learning in the presence of realistic, biased medical annotations. By distinguishing between low-level features and high-level semantics, it effectively utilizes all available data while protecting the model's core performance from annotation noise. Future work may extend this to multi-class segmentation and dynamic annotator behavior.
Summary updated