Because medical imaging data is growing at an impressive rate in hospitals and diagnostic centers, AI could radically transform hospitals and improve the accuracy of diagnoses. Still, since patient data must be kept safe and training data must comply with laws such as HIPAA and GDPR, it is difficult to use the traditional approach that brings together all the data for training. In recent times, Federated Learning (FL) has become a way of training AI using the power of multiple organizations without exchanging their raw data. This paper details federated learning approaches made for medical image analysis, with examples of classification and segmentation and addresses major issues about data privacy, the success of models and the system’s ability to scale. We study the effects of several FL methods and aggregation plans on different datasets collected at NIH and including a chest x-ray set and a tumor collection. Results from our study point out that models trained on a FL basis perform just as well as those trained with centralized methods and they still protect privacy because training data stays at the local sites. Other issues that slow down the use of FL in medicine include large shifts in data distribution, huge costs for communicating during training and the threat of attacks known as adversarial examples. We come up with solutions such as personalizing models, compressing gradients, using differential privacy and employing robust means for aggregation to deal with the described limitations. Model interpretability, secure multi-party computation and blockchain-backed audit trails are given special importance to ensure the system is ethical and trustworthy. According to this study, federated learning is a promising and responsible strategy to use AI in healthcare. To conclude, we propose advancing FL systems to be more robust, transparent and able to cooperate with other software which will support using them at scale in various medical imaging fields.
Kumar et al. (Mon,) studied this question.