Medical image segmentation plays a pivotal role in modern healthcare applications, significantly assisting healthcare professionals in making accurate diagnosis. However, limited annotated data hinders the development of robust medical image segmentation models due to strict patient privacy regulations. Semi-supervised learning offers a solution by leveraging both labeled and unlabeled data, enabling models to learn from larger datasets while capitalizing on limited labeled data. Despite its benefits, semi-supervised learning continues to face concerns related to data privacy. Federated learning (FL) addresses these concerns by enabling multiple parties to collaboratively train a model without sharing their local data. As the need for data-efficient and privacy-preserving machine learning techniques grows, semi-supervised federated learning emerges as a potential paradigm to tackle this issue. Our work underscores the importance of semi-supervised federated learning approaches that can effectively leverage both labeled and unlabeled data distributed across multiple sites. However, non-Independent and Identically Distributed (non-IID) data and class data imbalance remain significant challenges in semi-supervised federated learning, hindering overall performance. To address these issues, we propose the Federated Learning-Pseudo labeled Segmentation with Class balance (FL-PSeC) framework, which focuses on tackling class imbalance across clients’ local data by assigning higher weightage to underrepresented classes when generating annotations from unlabeled data. This strategy not only improves the IID nature of the data but also leads to enhanced global model performance. Our experiments on HAM10000, BUS, BUSIS, and UDIAT datasets demonstrate the efficacy of FL-PSeC in adapting to a class imbalance in FL, outperforming existing methods.