Nuclei segmentation is crucial for cancer diagnosis but faces high annotation costs due to dense nuclei distribution. Weakly supervised learning with point annotations alleviates this burden, yet single-center data are limited, and centralized datasets are hindered by privacy concerns. Federated Learning (FL) enables multi-institution collaboration while preserving privacy, but non-Independent and Identically Distributed (non-IID) data distribution—particularly style heterogeneity from staining and equipment variations—complicates model aggregation. In this paper, we propose Federated learning with Style Perturbation and Clustering (FedSPC), a novel framework that integrates a Federated Style Perturbation (FedSP) model and a Federated Style Clustering (FedSC) strategy. During training, FedSP applies style adversarial perturbation to extract and adapt local style features, reducing local style bias. Meanwhile, FedSC groups clients by style similarity and adjusts aggregation weights based on intra-group performance, mitigating fairness propagation bias. FedSPC overcomes pathological image style heterogeneity through the combined use of FedSP and FedSC, delivering a practical FL solution for medical imaging. Evaluated against existing federated weakly supervised frameworks, conventional methods, and aggregation schemes, our approach significantly outperforms alternatives in nuclei segmentation tasks. Experiments confirm FedSPC’s superiority in handling style diversity and improving segmentation accuracy under federated settings. Our code is available at https://github.com/Qyizos/FedSPC.

Retinal images play an essential role in the early diagnosis of ophthalmic diseases. Automatic segmentation of retinal vessels in color fundus images is challenging due to the morphological differences between the retinal vessels and the low-contrast background. At the same time, automated models struggle to capture representative and discriminative retinal vascular features. To fully utilize the structural information of the retinal blood vessels, we propose a novel deep learning network called Pre-Activated Convolution Residual and Triple Attention Mechanism Network (PCRTAM-Net). PCRTAM-Net uses the pre-activated dropout convolution residual method to improve the feature learning ability of the network. In addition, the residual atrous convolution spatial pyramid is integrated into both ends of the network encoder to extract multiscale information and improve blood vessel information flow. A triple attention mechanism is proposed to extract the structural information between vessel contexts and to learn long-range feature dependencies. We evaluate the proposed PCRTAM-Net on four publicly available datasets, DRIVE, CHASE_DB1, STARE, and HRF. Our model achieves state-of-the-art performance of 97.10%, 97.70%, 97.68%, and 97.14% for ACC and 83.05%, 82.26%, 84.64%, and 81.16% for F1, respectively.