Problem

The global healthcare sector faces challenges in accurate and timely lung disease diagnosis, such as COVID-19 and pneumonia, often exacerbated by the need for secure and distributed processing of sensitive medical imaging data. Traditional centralized AI models may struggle with data privacy concerns and the limited availability of large, diverse datasets across individual institutions, leading to issues like overfitting and suboptimal performance.

Method

The researchers propose a hybrid federated learning-enabled ensemble model for lung disease detection. The approach involves collecting and preprocessing X-ray image data, then training individual deep convolutional neural networks (VGG19, InceptionV3, DenseNet201) and a SWIN Transformer model. These trained models are ensembled to create a fusion model. This hybrid model is then integrated into a federated learning framework, where local models at different hospitals contribute updates to a central global model without sharing raw patient data, ensuring data security and privacy. The global model is iteratively refined by aggregating the best-performing local models.

Results

Individual models achieved notable validation accuracies: VGG19 at 94.4%, DenseNet201 at 94.1%, and Inception V3 at 94.5%. The SWIN Transformer model showed a validation accuracy of 82.5%. The final hybrid fusion model, combining these approaches, achieved a validation accuracy of 97.0% when summing weights and 94.0% when averaging weights. Training accuracy for the fusion model was approximately 99%, with a categorical validation accuracy of around 96%, indicating good performance but also a slight overfitting tendency. ROC-AUC curves demonstrated the fusion model's superior performance compared to individual models, and confusion matrix analysis indicated better true-positive predictions from the fusion model.

Implications

This hybrid federated learning approach offers a more accurate and secure method for lung disease diagnosis, potentially assisting medical practitioners in detecting COVID-19 and pneumonia from X-ray images. The integration of federated learning addresses critical data privacy concerns in healthcare. Future work includes addressing concept drift in federated learning, improving algorithm efficiency, performing statistical analysis of datasets, and exploring the application of this system to other disease detection processes. The authors emphasize that while technology assists medical treatment, it remains a supportive tool due to the inherent complexities and variances in diseases and treatment protocols.