VisionGraph-FPN: Graph Neural Network-Based Feature Pyramid Network for Robust Pneumonia Detection using Chest X-Ray Medical Image Dataset

Pneumonia is a lung infection caused by bacteria and viruses. If left untreated, it may lead to serious complications, and in some cases resulting in death. It is hence crucial to identify pneumonia through early diagnosis and start treatment in the initial stages rather than later. In this paper, the intricacies of subtle visual patterns and overlapping features in chest X-rays (CXRs) can be overcome. The proposed VisionGraph-FPN model replaces Faster R-CNN’s backbone architectures like ResNet and VGG with a custom feature pyramid network (FPN) for the purpose of object detection. This custom architecture incorporates Graph Neural Networks (GNN) to increase spatial awareness and contextual understanding. Integrating a U-Net architecture with residual connections, the aim is to preserve spatial details while leveraging deep feature extraction. In addition, the isotropic model in the left leg of the U-Net is used while successfully implementing the pyramid model in the reverse backbone. The resultant feature map obtained after the process is then passed into the FRCNN’s modules like the Region Proposal Network (RPN) and the Region of Interest (ROI) Pooling to predict the final bounding boxes. Notably, this approach achieves an Intersection over Union (IoU) accuracy of 85.71%, demonstrating its effectiveness in precise pneumonia detection. This model is trained using a medical image dataset, namely the RSNA Pneumonia Detection Challenge dataset.