Federated Transfer Learning for Energy Efficient Privacy-preserving Medical Image Classification

Abstract

The deep convolutional neural networks are widely used in medical image classification tasks. In some cases, they have outperformed physicians and achieved significant results. Unlike natural images, medical image dataset are very hard to collect, because they are protected by the privacy regulations to preserve patient's anonymity and requires a great deal of professional expertise to label them. However, because of the easier access and availability of high-performance computational resources, leveraging deep neural networks to detect diseases is becoming increasingly popular and common practice among healthcare researchers. As a result, considerable amount of energy is consumed to find an optimal and effective solution, which has a huge impact on our environment and contributes to global warming to some level.To address these challenges and reduce the carbon footprint caused by the deep learning practitioners, we attempted to combine the advantages of both federated learning and transfer learning for the medical image classification task in our study. Our findings suggest that federated transfer learning could be an useful technique to minimize computational costs and energy efficient, while maintaining privacy and addressing the problem of data scarcity. Moreover, this approach can be applied to solve other healthcare related tasks.