Artificial intelligence has shown remarkable ability to detect diseases from medical images. Models trained on large datasets of chest X-rays, MRI scans, and retinal photographs can match or exceed the diagnostic accuracy of experienced radiologists. However, there is a critical barrier preventing these AI systems from reaching their full potential: patient data privacy.
More than 30 percent of healthcare organizations globally have experienced a data breach in the past year alone. Centralizing patient data from multiple hospitals into a single training database, the traditional approach for building AI models, creates enormous privacy and legal risks. Sharing identifiable medical images across institutional borders is often prohibited by law, and the technical risks of a security breach are substantial.
The result is a fragmented landscape where AI models are trained in isolation on small, single-institution datasets, limiting their accuracy and generalizability. A model trained only at one hospital may perform poorly when deployed at another institution with different patient populations, imaging equipment, or clinical protocols.
Federated learning was developed to solve this problem. Instead of pooling data at a central location, federated learning lets each hospital train a model locally on its own data, then share only the model's learned parameters, not the raw data, with a central server that combines insights from all participating institutions. This preserves privacy while enabling collaborative learning.
This study tested a federated learning framework on three medical imaging tasks: classifying chest X-rays for tuberculosis detection, classifying brain MRI scans into tumor categories (no tumor, meningioma, glioma, pituitary tumor), and staging diabetic retinopathy from retinal photographs across five severity levels.
The core contribution of the study is a novel adaptive aggregation method that dynamically switches between two established approaches for combining model updates from multiple clients: Federated Averaging (FedAvg), which is computationally efficient but less effective when client data distributions differ widely, and Federated SGD (FedSGD), which is more accurate under high divergence but computationally more demanding.
The system measures the divergence between local models and the global model after each training round. When divergence is low, meaning all hospital models are learning similar things, FedAvg is used. When divergence is high, indicating that different hospitals are seeing very different patient populations, FedSGD takes over to ensure the updates are properly incorporated. A threshold value of 0.2 was found to provide the optimal balance between performance and efficiency.
Data was distributed across 10 simulated clients in a non-IID (non-identically distributed) manner, mimicking how real hospitals differ in disease prevalence, data volume, and imaging quality. This realistic simulation is critical for validating whether the framework would work in actual multi-hospital deployments.
Two well-established deep learning architectures were used as baseline models. VGG16 is a 16-layer convolutional network known for its depth and ability to capture detailed hierarchical patterns in images, making it effective for detecting subtle lesions or tumor boundaries. However, its large size (552 MB) makes it computationally expensive to share across federated clients.
GoogLeNet uses a modular design called inception modules, where the same image region is analyzed by filters of multiple different sizes simultaneously, allowing the model to detect both fine-grained details like microaneurysms in retinal images and large-scale structures like tumor margins. Its more compact size (27 MB) makes it far more efficient for federated deployment.
Two modern architectures were also evaluated to test the scalability of the framework. EfficientNetV2 uses compound scaling to simultaneously optimize the depth, width, and resolution of the network, achieving high accuracy with lower computational cost. ResNet-RS uses skip connections that allow information to bypass layers, enabling deeper networks to train more stably and extract complex hierarchical features.
All models were pre-trained on ImageNet, a large database of natural photographs, and then fine-tuned on the medical datasets. This transfer learning approach enables the models to leverage general visual knowledge while adapting to the specific patterns present in medical imaging, drastically reducing the amount of medical data needed for training.
In the centralized setting with differential privacy, models achieved the highest accuracy, ranging from 96.0 to 98.3 percent across datasets. This establishes the upper performance bound since centralized training uses all data together, but requires data sharing.
In the federated setting, the adaptive aggregation method consistently outperformed both FedAvg and FedSGD individually. For the diabetic retinopathy dataset, GoogLeNet achieved 96.3 percent accuracy with adaptive aggregation versus 95.9 percent with FedAvg. The improvements were larger for more heterogeneous datasets: approximately 1.5 percent improvement for TB X-rays, 2.3 percent for brain tumor MRI, and 3.0 percent for diabetic retinopathy.
Modern architectures EfficientNetV2 and ResNet-RS further elevated performance. EfficientNetV2 achieved 98.6 percent accuracy on brain tumor MRI and 98.3 percent on TB chest X-rays, exceeding the baseline models. It also handled imbalanced data distributions better, showing strong performance on rare disease classes like proliferative diabetic retinopathy.
The analysis of switching frequency revealed that the framework's behavior adapted appropriately to each dataset. For TB X-rays, where data is relatively similar across clients, FedAvg was used in 88 percent of rounds. For diabetic retinopathy, which has higher variability, FedSGD was triggered in 32 percent of rounds, demonstrating that the system correctly identified when finer gradient updates were needed.
The framework employs two complementary privacy techniques. In the centralized comparison baseline, differential privacy was used, which adds mathematically calibrated noise to model gradients during training. This provides a formal guarantee that no individual patient's data can be inferred from the trained model, even by an adversary with access to the model's parameters.
In the federated setting, privacy is preserved structurally: raw data never leaves each client's local server. Only model weights or gradient updates are transmitted to the central aggregation server. Even these transmitted updates reveal very little about individual patients, especially when combined from many clients.
The study showed that applying differential privacy in the centralized setting caused only a marginal reduction in accuracy compared to training without privacy protections. This demonstrates that strong privacy guarantees and high diagnostic accuracy are not mutually exclusive, which has been a concern preventing wider adoption of privacy-preserving AI in healthcare.
Communication overhead, the bandwidth required to share model updates across clients and the central server, varied significantly between architectures. GoogLeNet's 27 MB model size produced manageable overhead (270 MB per round), while VGG16's 552 MB model incurred a total of 276,000 MB across 50 rounds, highlighting a real practical concern for deploying larger models in resource-constrained hospital networks.
The practical significance of this research is enabling hospitals to contribute to powerful shared AI models without ever exposing their patients' medical records. A hospital in a rural area with limited patient volume could participate in training a globally accurate diagnostic model, and simultaneously benefit from the knowledge learned at large urban medical centers.
The non-IID data distribution used in the experiments realistically mirrors how different hospitals serve different patient populations. A hospital that predominantly treats elderly patients will have different imaging patterns than one serving younger demographics, and the adaptive aggregation method showed it can handle these differences without compromising global model quality.
The framework's generalizability across three very different types of medical imaging (chest X-rays, brain MRI, retinal photography) suggests it could extend to other imaging modalities relevant to cancer diagnosis. The same principles could apply to digitized pathology slides, CT scans for lung cancer screening, or mammography for breast cancer detection.
An important limitation acknowledged by the researchers is the computational overhead of larger models. While EfficientNetV2 achieved the best accuracy, its deployment across 10 federated clients still required substantial resources. Future work should focus on model compression techniques and more efficient communication protocols to make the framework practical for hospitals with limited infrastructure.
This study demonstrates that the combination of transfer learning and federated learning, enhanced by an adaptive aggregation mechanism, delivers accurate, privacy-preserving medical image classification that is scalable across multiple institutions and imaging types.
The novel adaptive aggregation approach, which dynamically selects the aggregation strategy based on real-time measurement of data divergence across clients, represents a meaningful advance over static federated learning methods. It addresses one of the fundamental challenges of real-world federated deployments, namely the inherent heterogeneity of data across medical institutions.
The consistent performance gains from modern architectures like EfficientNetV2 suggest that the framework will continue to improve as better neural network designs emerge, without requiring any fundamental changes to the federated learning infrastructure. This adaptability makes the framework future-proof.
Future research directions include extending the framework to additional medical imaging types, testing with larger numbers of federated clients, incorporating more sophisticated privacy guarantees, and developing better model compression strategies to reduce communication overhead. The ultimate goal is a healthcare ecosystem where AI models improve continuously as more institutions contribute, without any single institution ever having to compromise patient privacy.