MLFF-GNN: A Multi-Level Feature-Fusion Graph Neural Network for CXR-Image-Based Pneumonia Classification

Pneumonia does great harm to children's health. Although Graph Neural Network (GNN) has made great progress in the classification of pneumonia images, they usually use the permutation-invariant message-passing mechanism to capture the information between nodes, which is difficult to capture the interaction information between nodes. In addition, the quality of chest X-ray (CXR) images of pneumonia was uneven, and the classification model was affected by the redundant information in the pneumonia images, which could not effectively capture key information to classification. To address these problems, we proposed the multi-level feature-fusion graph neural network (MLFF-GNN) and a novel graph neural network block named the LSTM-enhanced graph network (LEGN) block. The LEGN block is designed to capture interaction information among nodes, including synergistic LEGN information within the interactions. Simultaneously, we introduce an information theory framework to explain the capture of interaction between nodes, including synergistic information within these interactions. In addition, we proposed a novel multi-level feature extraction (MLFE) module that utilizes a Convolutional Neural Network (CNN) pre-training model to extract multi-level feature maps from pneumonia CXR images and employs transfer learning to capture key features. On the two-category pneumonia classification dataset, MLFF-GNN surpasses 10 state-of-the-art (SOTA) models, with a notable 1.35% increase in accuracy. It achieves a sensitivity, precision, and accuracy of 99.30%, 99.30%, and 98.97%, respectively. On the two-category pneumonia classification dataset, MLFF-GNN maintains a high-performance standard with sensitivity, precision, and accuracy rates of 97.78%, demonstrating superior performance compared to existing models. The experimental outcomes demonstrate that the proposed model can extract the key features of CXR images of pneumonia, and can effectively capture the interaction information between nodes, which is superior to the existing methods. The proposed MLFF-GNN significantly aids clinicians by providing robust analytical support for the diagnosis of pneumonia from CXR images.