Federated semi-supervised calibrated efficient fine-tuning of foundation models for medical image classification

Federated learning provides a privacy-preserving paradigm for training models across distributed clients without sharing raw data. However, the scarcity of labeled data at individual clients poses a significant challenge to effective local model training. This work addresses the practical scenario where both labeled and unlabeled clients coexist, framing it as a federated sem-supervised learning (FSSL) problem. We introduce FedSemiCET, a federated semi-supervised calibrated efficient tuning framework for medical image classification using foundation models. Our framework consists of three main components: First, we develop a step-wise aggregation and smooth update strategy that mitigates the negative effects of client heterogeneity. Second, we design a multi-binary classification approach to assess - relationships; when integrated with conventional probability estimates, this yields more reliable confidence scores for unlabeled clients and reduces the influence of inaccurate pseudo-labels. Third, we employ a calibrated efficient tuning method with very few trainable parameters to decrease computational and communication costs in federated learning. Experimental evaluation on three medical image classification tasks demonstrates that our framework consistently outperforms current state-of-the-art methods while maintaining robust generalization capabilities.