Ulcerative Colitis is a chronic inflammatory bowel disease characterized by varying degrees of disease severity, often measured using the Mayo endoscopic score. The handling of medical data, encompassing sensitive personal health records and clinical information, necessitates stringent privacy protections. This study introduces a federated deep learning framework that leverages fine-tuned deep learning models trained via transfer learning on a dataset of endoscopic images. Employing DenseNet-121 as the foundational architecture, this approach enables the extraction and encoding of generic descriptors from ulcerative colitis images across multiple clients. We evaluated the training efficiency of our federated learning model in comparison to traditional centralized model, where the federated learning model outperformed the centralized approach by achieving an F1-score of 78% and a Quadratic Weighted Kappa score of 71%, and enhancing convergence speeds in terms of the reduction in the number of training rounds required. The results affirm the model’s capability in accurately diagnosing ulcerative colitis while ensuring the confidentiality of patient data, underscoring the viability of federated learning in sensitive healthcare applications.