Explainable Generative Attention Mechanisms for Chest X-Ray Medical Image Synthesis and Diagnosis of Pediatric Pneumonia

Pediatric pneumonia is a global health challenge currently addressed through medical imaging and analysis of chest X-rays to inform the diagnosis, therefore, previous research has focused exploring the use of deep learning models to improve pediatric pneumonia diagnosis. Unfortunately, there is still limited availability of good sufficient quality well-annotated balanced medical image datasets to train and evaluate the deep learning models. Most importantly, the deep learning models are “black-box" in nature hence hindering trustworthy their adoption and use among the physicians. In this study, we demonstrate the use of generative AI to synthesis pediatric chest X-ray images for deep learning model training and explainable AI techniques to improve the trust and adoption of such models in pediatric medical imaging and analysis. We address the dataset limitations using synthetic chest X-ray images generated using Generative Adversarial Networks (GANs) models with random noise to address mode collapse during training. With Transfer learning, we built Vision Transformers, Graph Convolutional Network (GCN), VGG16, and a custom CNNs with 85.0%, 99.6%, 79.3%, 75.0% accuracy at image classification respectively. Graph Attention Networks (GAT) were then used to learn the complex relationships and patterns within the Graph Convolutional Networks and interpretability is achieved using Generative Adversarial Saliency Maps, Gradient Attention Rollout, Intergrated Gradients and Grad-CAM.