Rajagopal, SmithaSmithaRajagopalShieh, Chin-ShiuhChin-ShiuhShiehShankar, Siva SSiva SShankarMaithili, KKMaithiliChakrabarti, PrasunPrasunChakrabarti2026-04-272026-04-272026http://dx.doi.org/10.1007/s00500-025-11025-9https://www.scopus.com/pages/publications/105029629816https://gnanaganga.alliance.edu.in/handle/123456789/10709Currently, artificial intelligence (AI) has been widely used in medical imaging for accurate disease classification. However, the current existing approaches offer centralized learning environments that make them prone to security risks. These security risks may hinder the accuracy of disease classification, leading to a reduced survival rate. To address these issues, a novel multi-class brain tumor classification strategy was proposed in this work using federated learning. The proposed approach integrates the features of federated learning with the Deep CapsNet approach (FL-DCN), offering both accuracy and privacy. The proposed framework consists of four modules: data collection and preprocessing, feature extraction, and classification. In the data collection and preprocessing module, the brain Magnetic Resonance Imaging (MRI) images are collected and preprocessed to ensure consistency and quality. In the feature extraction module, a U-Net architecture was employed for extracting the most significant attributes present in the images. Here, a dual outcome structure was enabled to implement the segmentation and feature extraction process simultaneously. The outcome of this stage is high-level feature and tumor masks for CapsNet. In the classification module, a federated learning environment was created, and the extracted features are distributed across multiple local nodes. Here, each node has trained three elements, such as primary capsules, digit capsules, and dynamic routing. Moreover, primary capsules contain 32 capsules and 8D activation vectors; similarly, digit capsules consist of 4 capsules and 16D vectors used to match the tumor classes. Consequently, the local node uses Deep CapsNet to learn the spatial and hierarchical patterns within the images. Finally, Federated Averaging (FedAvg) was employed to combine the model parameters of the local models into a global model, which produces the classification outcomes. The presented methodology is implemented in the Python tool and validated across the public database. The implementation results are determined and validated across the existing technologies in terms of accuracy, precision, recall, F-measure, error rate, and computational efficiency. While comparing the conventional model, this study demonstrates that the FL framework can attain privacy-protected brain tumor classification from brain MRI image datasets without compromising much accuracy as well as precision parameters. © The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2026.enBrain Tumor ClassificationDeep Capsule Neural NetworkDisease PredictionFederated LearningArticle