Pixel-wise Divide and Conquer for Federated Vessel Segmentation

Tian Chen, Wenke Huang, Zhihao Wang, Zekun Shi, He Li, Wenhui Dong, Mang Ye, Bo Du, Yongchao Xu

Abstract

Vessel segmentation is a critical task in medical imaging, yet it faces significant hurdles in Federated Learning (FL) environments due to varying vessel morphologies across institutions and severe class imbalance. We introduce Federated Vessel-Aware Calibration (FVAC), a "divide and conquer" approach designed to address these challenges. FVAC employs pixel-wise uncertainty to guide models toward difficult structural features and decouples foreground-background alignment to prevent semantic drift. Our results demonstrate substantial improvements in segmentation accuracy, particularly for fine-grained vessels.

Introduction

Accurate vessel segmentation is essential for diagnosing conditions such as glaucoma and coronary artery disease. While Federated Learning (FL) enables multi-institutional collaboration without compromising data privacy, it encounters two primary obstacles in this domain:

  1. Morphological Inconsistency: Variations in vessel thickness and complexity across different clinical datasets cause standard FL models to favor dominant features, often neglecting subtle but clinically significant fine vessels.
  2. Vessel-Background Imbalance: Blood vessels typically occupy a minute fraction of medical images. Traditional feature alignment in FL often over-optimizes for the background, leading to "semantic drift" where the representation of the target vessels becomes unstable.

Methodology: Federated Vessel-Aware Calibration (FVAC)

The proposed FVAC framework addresses these issues through two innovative modules: Federated Morphology Uncertainty Guidance (FMUG) and Federated Vessel Decoupling Alignment (FVDA).

Overview of the FVAC framework, illustrating the general architecture (A), the Federated Morphology Uncertainty Guidance (B), and the Federated Vessel Decoupling Alignment (C).

Federated Morphology Uncertainty Guidance (FMUG)

FMUG shifts the focus from global image statistics to "hard kernels"—specific pixels where the model is prone to failure, such as blurry edges or complex intersections. By synthesizing global uncertainty from the server and local uncertainty from the client, FMUG generates a pixel-wise weight map. This map informs a weighted cross-entropy loss, compelling the model to prioritize these challenging regions. This synergy prevents local overfitting and improves generalization across diverse vessel types.

Federated Vessel Decoupling Alignment (FVDA)

To combat the imbalance-driven semantic drift, FVDA moves away from holistic feature map alignment. Instead, it utilizes ground-truth labels to decouple features into foreground (vessels) and background components. These components are aligned independently between local and global models. This ensures that vessel-specific features are synchronized effectively across the federated network without being overwhelmed by the background majority.

Experimental Setup

The FVAC framework was evaluated across two imaging modalities:

  • Color Fundus Images: Datasets included DRIVE, STARE, and CHASEDB1.
  • Optical Coherence Tomography Angiography (OCTA): Datasets included ROSE-1 (SVC and SD subsets) and OCT-500 (3MM and 6MM subsets). OCTA data is notably challenging due to its dense and intricate vessel networks.

Key Results and Discussion

FVAC consistently outperformed state-of-the-art FL methods, including FedAvg, FedProx, and MOON.

  • Quantitative Performance: On the complex OCTA-6MM dataset, FVAC achieved a 6.44% improvement in the Dice coefficient over the FedAvg baseline.
  • Qualitative Superiority: Visual analysis confirms that FVAC excels at maintaining the continuity of thin capillaries and accurately resolving complex intersections, areas where baseline models frequently fail.
  • Robustness: The framework demonstrated high resilience to data heterogeneity across different simulated hospital clients.

Limitations and Future Work

Despite its effectiveness, FVAC introduces some computational overhead due to the additional forward pass required for global uncertainty calculation. Furthermore, the reliance on global model guidance presents potential security risks if the server is compromised. Future research will focus on optimizing computational efficiency and developing robust defensive mechanisms against adversarial interference.

Summary updated

Abstract

Accurate vessel segmentation is essential for diagnosing and managing vascular and ophthalmic diseases. Traditional learning-based vessel segmentation methods heavily rely on high-quality, pixel-level annotated datasets. However, segmentation performance suffers significantly when applied in federated learning settings due to vessel morphology inconsistency and vessel-background imbalance. The former limits the ability of models to capture fine-grained vessels, while the latter overemphasizes background pixels and biases the model towards them. To address these challenges, we propose a novel method named Federated Vessel-Aware Calibration (FVAC), which leverages global uncertainty to provide differentiated guidance for clients, focusing on pixels of various morphologies that are difficult to distinguish. Furthermore, we introduce a foreground-background decoupling alignment strategy that utilizes more stable and balanced global features to mitigate semantic drift caused by vessel-background imbalance in local clients. Comprehensive experiments confirm the effectiveness of our method
BibTeX
@inproceedings{chen2025pixel,
  author = {Tian Chen and Wenke Huang and Zhihao Wang and Zekun Shi and He Li and Wenhui Dong and Mang Ye and Bo Du and Yongchao Xu},
  title = {{Pixel-wise Divide and Conquer for Federated Vessel Segmentation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/540},
}
Figure