Abstract
Federated Learning (FL) offers a promising path for collaborative medical AI by enabling multi-institutional training without the exchange of sensitive patient data. However, real-world clinical environments are characterized by both statistical heterogeneity (Non-IID data) and model heterogeneity (diverse hardware and architectural preferences). Current solutions for model heterogeneity often rely on shared public datasets for knowledge distillation, which are frequently unavailable or biased in medical contexts. This paper discusses MH-pFLGB, a framework that introduces a "global bypass" mechanism to bridge the heterogeneity gap without external data dependencies. By utilizing a lightweight messenger model and a novel feature-fusion module, MH-pFLGB achieves state-of-the-art performance in medical image classification and segmentation while maintaining high computational efficiency.
1. Introduction
The deployment of Federated Learning in healthcare faces two primary obstacles. First, statistical heterogeneity arises because different hospitals serve different patient demographics and clinical specialties, leading to non-identically and independently distributed (Non-IID) data. Second, model heterogeneity occurs when participating institutions utilize different neural network architectures (e.g., ResNet, MobileNet, or VGG) due to varying local hardware constraints or legacy software requirements.
Existing approaches to model-heterogeneous FL typically employ Knowledge Distillation (KD). These methods require a central "public dataset" to align the outputs of diverse local models. In the medical field, obtaining a high-quality, privacy-cleared public dataset that aligns with the specific local task is often unfeasible. MH-pFLGB addresses these challenges by facilitating knowledge transfer through a shared, lightweight "bypass" model that operates alongside local architectures, eliminating the need for a public proxy dataset.
2. Methodology
The MH-pFLGB framework consists of three core components designed to facilitate collaboration across diverse architectures:
2.1 The Local Model and Global Bypass
Each client in the network maintains its own unique Local Model, consisting of a feature extractor (body) and a classifier/segmenter (head). Parallel to this, every client hosts a copy of the Global Bypass Model. This bypass is a standardized, lightweight "messenger" model shared across the entire federated network. Unlike the local models, only the parameters of the global bypass are aggregated and synchronized by the central server.
2.2 Features Weighted Fusion
The "secret sauce" of the framework is the Features Weighted Fusion module. Rather than simply averaging the predictions of the local and bypass models, MH-pFLGB performs fusion at the feature level. A Softmax-based gating mechanism learns to dynamically weigh the features extracted by the local model against those from the global bypass. This allows the system to decide, on an input-by-input basis, which model's representation is more reliable for the specific task.
2.3 The Two-Stage Training Pipeline
To ensure effective knowledge transfer, training is conducted in two alternating stages:
- Local Stage: The global bypass parameters are frozen. The local model is trained on local data, utilizing the insights embedded within the bypass to refine its own feature extraction.
- Global Stage: The local model is frozen. The global bypass is fine-tuned on the local data to "absorb" the unique characteristics of the client's dataset. These updated bypass parameters are then sent to the server for global aggregation.
3. Experimental Setup
The researchers evaluated MH-pFLGB across three diverse medical imaging tasks to test its versatility:
- BreaKHis: A histopathological database for breast cancer, focusing on multi-resolution classification.
- OCT2017: A grayscale dataset for Optical Coherence Tomography, targeting eye disease classification.
- Polyp Segmentation: A composite dataset (ColonDB, ETIS, ClinicDB, and Kvasir-SEG) designed to test the model's robustness across different surgical imaging environments and camera types.
4. Results and Discussion
4.1 Performance Benchmarks
MH-pFLGB consistently outperformed both standard Federated Learning (FedAvg) and specialized heterogeneous methods (FedMD, KT-pFL). In the BreaKHis classification task, MH-pFLGB achieved an average accuracy of 0.8597, a significant improvement over the 0.7925 achieved by the next best heterogeneous method, KT-pFL. In the complex task of polyp segmentation, the framework reached a Dice score of 0.7693, substantially higher than the 0.6897 achieved by local training alone.
4.2 Computational Efficiency
A key advantage of the "messenger" approach is its efficiency. The global bypass is designed to be extremely small. In experimental trials, the bypass model often contained 100x to 1000x fewer parameters than the local models (e.g., 0.035M parameters compared to VGG's 40M). This ensures that the communication overhead and the computational "tax" for participating in the federated network remain negligible for the clients.
4.3 Ablation Studies
Ablation experiments confirmed that every component of the architecture is vital. Removing the global head, the global body, or the fusion module resulted in a measurable performance decline, validating the hypothesis that the bypass architecture is the primary driver of the framework's success.
5. Conclusion and Future Work
MH-pFLGB represents a significant step forward for personalized Federated Learning in medicine. By moving knowledge transfer from the prediction level to the feature level and utilizing a lightweight bypass, it removes the dependency on public datasets while accommodating diverse local architectures.
Future research may focus on reducing the number of communication rounds required (currently 100) and testing the framework's performance when bridging fundamentally different architectures, such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). Additionally, exploring a task-agnostic "foundation" bypass could further generalize the framework for multi-task medical applications.
Summary updated