COVID-19 detection using federated machine learning.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• The COVID-19 pandemic necessitates advanced tools for accurate case classification and prediction.
• Traditional machine learning models for COVID-19 analysis raise patient privacy concerns due to data access requirements.
• Federated learning emerges as a privacy-preserving alternative for decentralized model training.

Purpose of the Study:

• To compare the efficacy of federated learning against traditional machine learning for COVID-19 classification using chest X-ray images.
• To investigate the impact of various factors (activation function, optimizer, learning rate, rounds, data size) on model performance.
• To evaluate prediction accuracy, loss, and performance speed of both federated and traditional models.

Main Methods:

• Developed and trained two machine learning models: one using federated learning and another using traditional learning, with Keras and TensorFlow Federated.
• Utilized a dataset comprising descriptive data and chest X-ray (CXR) images from COVID-19 patients.
• Systematically analyzed the influence of activation functions (softmax), optimizers (SGD), learning rate, training rounds, and data size on model accuracy and loss.

Main Results:

• The softmax activation function and SGD optimizer yielded superior prediction accuracy and reduced loss in both model types.
• Adjusting the number of training rounds and learning rate had a minor impact on prediction accuracy and loss.
• Increasing dataset size did not significantly affect model prediction accuracy or loss.
• The federated learning model demonstrated higher prediction accuracy and lower loss than the traditional model, albeit with a longer performance time.

Conclusions:

• Federated learning provides a viable and effective approach for COVID-19 classification from chest X-rays, balancing accuracy with enhanced data privacy.
• Model performance is significantly influenced by the choice of activation function and optimizer, with softmax and SGD being optimal.
• While federated learning may incur higher computational time, its privacy-preserving benefits are crucial for handling sensitive patient data in pandemic scenarios.