Abstract
PathFL is a novel federated learning (FL) framework designed to tackle the extreme data heterogeneity found in multi-center pathology image segmentation. By implementing a "triple alignment" strategy—synchronizing data at the image, feature, and model aggregation levels—it achieves state-of-the-art performance across diverse scanners, organs, and imaging modalities while strictly preserving patient privacy.
Introduction
Pathology image segmentation, which involves identifying nuclei, cells, and tissues, is a cornerstone of modern cancer diagnosis. Training robust AI models for these tasks requires massive amounts of annotated data. However, pooling data from multiple hospitals is often prohibited by strict privacy regulations like HIPAA or GDPR, as well as the proprietary nature of clinical datasets.
Federated Learning (FL) offers a solution by allowing institutions to train a shared model without exchanging raw images. Despite this, pathology data is notoriously heterogeneous. Images vary significantly due to differing staining protocols, scanning equipment, and the specific organs under study. Standard FL algorithms, such as FedAvg, often struggle in these "non-IID" (non-independent and identically distributed) scenarios, as the global model tends to bias toward dominant data sources, leading to poor generalization on unique local datasets.
Methodology
PathFL introduces three specific alignment modules to harmonize the learning process across diverse clients:
- Collaborative Style Enhancement (CSE) - Image Level: This module facilitates the exchange of stylistic statistical information (mean and variance of pixel intensities) between clients. By applying these "styles" to local images using a random masking technique, each client's model is exposed to a wider variety of visual distributions without ever seeing another client's raw data.
- Adaptive Feature Alignment (AFA) - Feature Level: While images may look different, underlying biological structures remain similar. AFA ensures that the "latent" features extracted by models are aligned in a shared mathematical space, infusing local features with global statistical insights to promote consistency.
- Stratified Similarity Aggregation (SSA) - Model Level: Rather than simple weight averaging, the server uses Gaussian-distributed synthetic data to probe layer-wise similarity between clients. It assigns higher weights to layers showing stronger cross-client similarity, effectively filtering out noise and irrelevant local variations.
The framework is grounded in Vicinal Risk Minimization (VRM) theory, suggesting that models generalize better when trained on a distribution that includes "neighbors" of actual data points.
Experimental Setup and Datasets
The study evaluates PathFL on a massive scale, utilizing 16 distinct datasets organized into four primary heterogeneity scenarios:
- Cross-source: 7 independent datasets (e.g., CoNSeP, CRAG, TNBC) with different origins and disease types.
- Cross-modality: 3 modalities (Cytology, Fluorescence, Histology) from the ClusterSeg dataset.
- Cross-organ: 3 organs (Colorectum, Pancreas, Stomach) from the COSAS2024 challenge.
- Cross-scanner: 3 different Whole Slide Image (WSI) scanners (3d-1000, Kfbio-400, Teksqray-600p).
Key Results
PathFL consistently outperformed standard FL baselines (FedAvg, FedProx) and specialized methods (FedBN, HarmoFL) across all test scenarios:
- Cross-source: Achieved an average Dice score of 84.27%, a 1.65% improvement over FedAvg.
- Cross-modality: Observed a significant jump of 6.13% in Dice score compared to the baseline.
- Cross-organ & Cross-scanner: Improvements of 7.94% and 6.35% respectively.
- Foundation Model Compatibility: When tested with pathology-specific foundation models like CONCH and CHIEF, PathFL bridged the gap between generalized feature extraction and local adaptability, yielding the best overall results.
Discussion and Future Work
PathFL demonstrates that a holistic approach—aligning input, internal representation, and final aggregation—is significantly more effective than single-stage alignment. By decoupling "style" (color/texture) from "morphology" (shape/structure), the framework simulates a global dataset locally while maintaining privacy through synthetic data probing.
While the method introduces some communication overhead due to the exchange of style statistics and similarity scores, it represents a major step forward for multi-center medical AI. Future research will focus on reducing these communication costs and extending PathFL's adaptability to other medical imaging domains such as MRI or CT scans.
Summary updated