FedProx-Enhanced Federated Transfer Learning for Heterogeneous 3D Medical Image Classification

Abstract

Artificial intelligence has revolutionized the analysis of medical images, but the creation of robust models is still problematic because of the strict regulation of data privacy and the very nature of medical imaging data heterogeneity. Federated learning (FL) is a privacy-preserving method with convergence instability and worse performance, particularly with non-Independent and Identically Distributed (non-IID) data. In addressing these disadvantages, this paper suggests a new FL framework based on the integration of transfer learning (TL) and FedProx to increase the resilience and stability of 3D medical image classification models in the distributed hospital context. TL offers a knowledgeable starting point to local models that enhances adaptation to different client data, whereas FedProx presents a proximal term to decrease the effects of data variability. The proposed framework was tested on three 3D medical imaging datasets, namely, OrganMNIST, FractureMNIST, and NoduleMNIST, with non-IID data distributions of 10 clients. It has been experimentally demonstrated that the performance is significantly improved, with an accuracy improvement of up to 18.2 and higher precision, recall, and F1 scores than current FL methods. All in all, the suggested solution offers a good and privacy-aware solution to collaborative learning in heterogeneous 3D medical imaging settings.



Received: 4 August 2025 | Revised: 4 Novemeber 2025 | Accepted: 16 January 2026



Conflicts of Interest

The authors declare that there are no conflicts of interest to this work.



Data Availability Statement

The OrganMNIST, FractureMNIST, and NoduleMNIST datasets used in this study are part of the MedMNIST collection, a standardized benchmark suite of biomedical images for machine learning tasks. The data that support the findings of this study are openly available at https://medmnist.com/, https://github.com/MedMNIST/experiments/tree/main/MedMNIST3D, and https://github.com/TsingZ0/PFLlib/tree/master/dataset.



Author Contribution Statement

Manjunath Naganna: Conceptualization, Methodology, Software, Resources, Writing – original draft, Writing – review & editing, Visualization, Project administration. Guru Ramachandra Nayaka: Formal analysis, Supervision. Natesh Mahadev: Validation, Investigation, Resources, Data curation. Mayura Tapkire: Conceptualization, Validation, Data curation, Writing – review & editing.