Journal Article · 2025
Federated Few-Shot Continuous Learning for Medical Image Processing with Federated Feature Augmentation
Authors
Abstract
Developing robust models for medical image classification remains a critical challenge due to limited annotated data, patient privacy regulations, and the requirement for continuous adaptation in dynamic clinical environments. This study proposed an integrated learning system that combined Federated Learning, Few-Shot Learning, Federated Feature Augmentation, and Continuous Learning to address these constraints in malaria diagnosis. The federated setup enabled distributed healthcare institutions to collaboratively train deep learning models without centralizing patient-level data. The Few-Shot Learning module improved adaptability to localized data scarcity, while Federated Feature Augmentation facilitated cross-node generalization through anonymized latent feature exchange. A Continuous Learning mechanism allowed the system to incorporate evolving diagnostic patterns without performance degradation over time. The system was evaluated using the Thick Blood Smears dataset in a decentralized training environment. Quantitative results demonstrated that the Xception model achieved the highest overall performance, with 95.16% accuracy, 97.13% precision, 94.09% recall, 95.58% F1-score, and 98.02% area under the receiver operating characteristic curve. Additional experiments showed that the convolutional neural network model improved consistently across ten rounds of federated training, peaking at 90.92% accuracy and 95.29% AUC. Model explainability was ensured through post-hoc techniques including Gradient-weighted Class Activation Mapping and Local Interpretable Model-Agnostic Explanations, both of which confirmed that predictions were based on clinically relevant morphological features. The complete system was implemented using TensorFlow, OpenCV, Flower, Django, and Flutter to support federated orchestration, real-time mobile diagnostics, and secure deployment. The results validated the effectiveness of the proposed system in achieving high diagnostic accuracy, privacy preservation, and interpretability under data-constrained conditions. These findings highlight the system’s applicability for real-world deployment in decentralized healthcare settings such as those in Uganda.
Keywords
Research areas
Join Us
Ready to join MRU?
Applications are open for the next intake. Apply on the E-Portal in under twenty minutes, or talk to the admissions team first — whichever helps you more.