Federated Learning Model Using CBDPL for Medical Image Segmentation

Federated learning (FL) allows hospitals and medical centers to train models without sharing patient data. Instead of sending data to a central server, each institution trains the model locally and only shares updates, keeping patient information private. Our study proposes a novel methodology for medical image segmentation that leverages semi-supervised learning and confidence-based dynamic pseudo-labeling (CBDPL) to enhance model performance with limited labeled data. Each participating client site independently performs feature extraction and basic segmentation tasks. For labeled data, segmentation is guided by a ground-truth comparison and loss minimization. For unlabeled data, semi-supervised learning is applied, with a dynamic pseudo-labeling module that assigns pseudo-labels based on confidence scores, integrating only high-confidence predictions to refine the model iteratively. Client updates are then combined on a central server. It helps to improve the overall model while keeping each client’s data private. This approach demonstrates the potential for robust, privacy-preserving segmentation in multi-dataset medical image analysis settings, addressing labeled data scarcity and privacy concerns.