ImageGCN: Multi-Relational Image Graph Convolutional Networks for Disease Identification With Chest X-Rays

Area of Science:

• Computer Vision
• Medical Imaging
• Graph Neural Networks

Background:

• Existing image representation methods often process images independently, neglecting valuable inter-image relationships.
• Modeling these relationships can enhance image understanding, ensure model consistency, and improve explainability.

Purpose of the Study:

• To propose ImageGCN, a novel framework for inductive multi-relational image modeling using graph convolutional networks.
• To leverage image-level relations for more informative image representations, particularly in medical imaging applications.

Main Methods:

• ImageGCN is an end-to-end graph convolutional network framework that integrates pixel features with relational information.
• The model learns image representations by considering both intrinsic image properties and their connections to other images.

Main Results:

• ImageGCN demonstrated superior performance in disease identification and localization tasks on three open-source chest X-ray datasets (ChestX-ray14, CheXpert, MIMIC-CXR).
• The framework achieved comparable or better results than state-of-the-art methods, showcasing its effectiveness.

Conclusions:

• ImageGCN provides a powerful approach for learning informative image representations by incorporating relational information.
• The framework shows significant potential for improving diagnostic accuracy and efficiency in medical image analysis.