Artificial intelligence models for medical imaging require large, diverse training datasets to work well across different hospitals and patient populations. However, healthcare data is highly sensitive, and privacy regulations often prevent hospitals from sharing patient records or pathological images with other institutions.
Training an AI model on data from only one hospital creates problems: the model may become overfitted (it learns the quirks of that one hospital rather than general patterns), or it may produce biased results that do not apply to patients elsewhere. The more hospitals contribute, the more robust and generalizable the model becomes.
Federated learning was developed to solve this dilemma. Instead of sending raw patient data to a central server, each hospital trains a local version of the AI model on its own data. Only the model's learned parameters (its "knowledge," not the data itself) are shared and combined. This allows multi-institutional collaboration without any patient data ever leaving its hospital of origin.
While federated learning is promising in theory, its real-world performance can be affected by data heterogeneity - the fact that pathological images from different hospitals look different due to variations in staining protocols, scanners, and preparation techniques. This paper directly tests whether federated learning can match centralized learning in practice for endometrial cancer pathology.
The researchers collected hematoxylin and eosin (H&E)-stained whole slide images (WSIs) of endometrial cancer tissue from four South Korean hospitals. H&E staining is the standard technique in pathology for making tissue structures visible under a microscope; WSIs are high-resolution digital scans of entire glass slides.
Three hospitals (Sinchon Severance, Gangnam Severance, and Gachon University Gil Hospital) provided training data - standardized to 66 WSIs each to ensure balanced contributions. A fourth hospital (Kangdong Sacred Heart Hospital) provided an independent validation set of 14 WSIs from six patients that was never seen during training. Each cancer region was manually annotated by a pathologist to serve as the ground truth for training and evaluation.
The AI model used for tissue segmentation was U-Net, a deep learning architecture specifically designed and widely used for medical image segmentation. It learns to classify each pixel as either cancer or non-cancer tissue. The team applied three enhancement techniques: data augmentation (flipping images to create more training examples), color normalization using a method called Vahadane (to correct for hospital-specific staining differences), and a specialized FedYogi optimizer for the federated learning version (an improved algorithm for combining model updates from multiple institutions).
Performance was measured using three metrics: precision (of all areas the model flagged as cancer, what fraction actually was cancer), recall (of all actual cancer areas, what fraction did the model correctly identify), and Dice Similarity Coefficient (DSC) (an overall measure of how well predicted and true cancer regions overlap). In medical diagnosis, recall is particularly critical because missing cancer (a false negative) is generally more dangerous than a false alarm.
When comparing the final optimized models, federated learning achieved lower precision (76.3% vs. 79.3%) but higher recall (81.7% vs. 74.1%) and a higher DSC (78.5% vs. 75.9%) compared to centralized learning. The differences in precision and recall were statistically significant (p less than 0.05), while the DSC difference showed a trend toward significance (p = 0.06).
In practical terms, the federated learning model was better at catching cancer - it missed fewer real cancer regions (higher recall) - at the cost of occasionally flagging some non-cancer tissue as cancer (lower precision). Given that missing cancer in clinical settings has more severe consequences than a false alarm, the federated model's profile is arguably more clinically appropriate.
False-positive and false-negative rates were not significantly different between the two models (p = 0.79 and p = 0.77 respectively), suggesting that despite the precision-recall trade-off, the overall error distributions were comparable. The Bland-Altman analysis confirmed that performance differences were small and consistent across cases.
Data augmentation (simply flipping images horizontally and vertically) had the largest individual impact, boosting precision by about 3.7%, recall by 14.9%, and DSC by 11.5% - all highly significant. This shows that even simple augmentation can dramatically improve model performance when training data is limited at each hospital site.
Color normalization (adjusting for staining variation across hospitals) reduced precision by about 4.3% but improved recall by 2.3% and DSC by 3.5%. The trade-off favors recall - the more important clinical metric - suggesting that standardizing staining appearance helps the model generalize better to new hospitals.
Switching from the standard FedAvg optimizer (which simply averages model weights from all hospitals) to FedYogi (which handles heterogeneous data more intelligently) produced the most dramatic gains: precision improved by 9.8%, recall by 12.3%, and DSC by 10.8%. This highlights that the choice of aggregation algorithm is critical in federated learning with diverse hospital data.
This study demonstrates that federated learning can achieve comparable or superior performance to centralized learning for segmenting cancer tissue in endometrial cancer pathology images - without requiring any raw patient data to be shared across institutions. This directly addresses the core privacy challenge that has hampered multi-institutional medical AI development.
The practical implication is that hospitals and cancer centers could collaboratively develop better AI diagnostic tools while keeping all patient data on-site, complying with privacy regulations, and maintaining institutional data sovereignty. As training datasets grow to include more hospitals, federated models are expected to continue improving.
The authors note that challenges remain, including the need for careful selection of preprocessing steps, optimizer algorithms, and managing quality differences in data across institutions. Future work could explore applying federated learning to other cancer types, using larger multi-national datasets, and testing on even more diverse scanner and staining configurations to further validate robustness.