Intelligent Medical Image Analysis Using Secure and Explainable Federated Deep Learning

Abstract

AI has made a significant impact on healthcare applications, such as intelligent disease diagnosis and medical image analysis. However, centralized DL approaches introduce an concerns about privacy, security of patients and the transparency of how predictions are made, i.e., in a black box way. Federated Learning works well for the concept of many collaborating parties training a common model on their own sensitive data without having to share the actual medical records, so that difference seems rather significant. Explainable Artificial Intelligence (XAI) further increases human comprehension of deep learning predictions and makes it less a black box. In these paper, we propose a interpretable federated deep learning model for privacy-preserving medical image analysis in I-healthcare ecosystem. The TL; DR is convolutional neural networks + federated learning, with Grad-CAM based explainability - so you get privacy-preserving behaviour with some level of interpretability in the resulting healthcare intelligence. In the experiments, the evaluation shows that better classification accuracy and higher clinician trust than traditional centralized methods.  The proposed model has great potential to support future intelligent healthcare systems through building scalable, secure, and transparent AI-enabled diagnostic solutions.