Understanding the Impact of Client Heterogeneity on Ordinal Classification in Federated Medical Image Analysis

Deep learning methods have shown remarkable success in medical image classification, aiding in early disease detection and treatment. Many of these tasks, such as cancer staging or risk stratification, exhibit an inherent ordinal structure; however, existing solutions often reduce them to binary or purely nominal classifications, ignoring the valuable ordering information. Simultaneously, privacy and regulatory concerns have spurred the adoption of Federated Learning (FL), enabling collaborative model training without centralising sensitive patient data. Yet, FL in real-world medical scenarios faces significant challenges arising from heterogeneous client data, particularly when institutions differ widely in case severity or label distribution. In this work, we conduct the first in-depth study of Federated Ordinal Learning (FOL), introducing ordinal classification paradigms into FL pipelines and systematically evaluating their performance under increasing levels of data heterogeneity. We assess the benefits of ordinal classification within four FL frameworks: standard Federated Averaging (FedAvg) and three heterogeneity-focused approaches (FedProx, MOON, and FedALA). Our experiments reveal that ordinal methods can effectively maintain class ordering information even when institutional data exhibit severe imbalance or missing classes, offering valuable insights for developing robust, privacy-preserving AI systems in medical imaging. However, ordinal approaches still suffer from performance degradation in highly heterogeneous FL settings, underscoring the need for dedicated research on FL methods that explicitly account for ordinality.