Pathology characteristics-aware federated learning for weakly supervised nuclei segmentation

Abstract

In federated learning, a key challenge in nuclei segmentation lies in data heterogeneity, primarily resulting from the diverse sources of pathology images. Nuclei in pathological images are typically small and densely distributed, making accurate annotation highly labor-intensive and reliant on specialized expertise. Weakly supervised learning is widely adopted for this task, as it only requires point annotations at the centers of nuclei. However, point annotations lack precise boundary information, thereby exacerbating the difficulties introduced by data heterogeneity. To address this issue, we propose a preprocessing strategy that leverages the unique optical properties of H&E stained images to generate Contrast-Difference Enhanced Images (CDEI). These CDEI highlight nucleus boundaries to varying extents based on the tonal characteristics of H&E stained images. Building on this strategy, we propose a Multi-source Hierarchical Enhancement Network (MHEN) as the client-side architecture. MHEN takes both the H&E stained images and the corresponding CDEI as input, effectively mitigating the limitations of weak labels by incorporating enhanced boundary cues. Furthermore, considering the characteristics of nuclei segmentation, we design a Federated Nuclei-Weighted Aggregation strategy on the server side. This strategy estimates each client’s contribution to the global model by quantifying the number of nuclei present in its local pathology images. To thoroughly assess the effectiveness of our approach, we compare it with both conventional weakly supervised methods and existing federated weak supervision frameworks. The experimental results demonstrate that our proposed federated learning framework for weakly supervised nuclei segmentation significantly outperforms existing methods. Our source code is available on GitHub.¹