The integration of Artificial Intelligence (AI) into the medical field is revolutionizing healthcare by enabling efficient diagnostics and decision-making. A key challenge in medical AI is to ensure accurate predictions while maintaining data privacy and confidentiality. Several works have tried to address this challenge using Federated Learning (FL), but most of them are limited to simulated environments with no real-world implementation. We propose a real-time Websocket communication-based FL method for classifying medical images, which enables several hospitals to work together to train a deep learning (DL) model without disclosing private patient information. Each hospital, acting as a worker node, trains a local DL model on medical image datasets. Locally trained models generate weight parameters that are encrypted using homomorphic encryption (HE). These encrypted weights are then safely transmitted to a centralized master node, where they are aggregated to create an updated global DL model. The final model is then deployed back to the worker nodes, ensuring improved accuracy and generalization. Communications between the worker nodes and the master node are facilitated using WebSocket-based connections. Using FL and encryption techniques, this real-time implementation enhances model performance while preserving patient data confidentiality, demonstrating the practical feasibility of AI-driven secure healthcare solutions.