Label Co-Occurrence Learning With Graph Convolutional Networks for Multi-Label Chest X-Ray Image Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Graph Neural Networks

Background:

• Multi-label medical image analysis often overlooks complex relationships between pathologies.
• Existing methods primarily focus on direct label prediction, neglecting valuable graph-structured information.

Purpose of the Study:

• To propose CheXGCN, a novel framework for multi-label chest X-ray (CXR) image classification.
• To explicitly model and utilize the dependencies and co-occurrence patterns among pathologies.

Main Methods:

• Developed a framework integrating an Image Feature Embedding (IFE) module and a Label Co-occurrence Learning (LCL) module.
• Employed Graph Convolutional Networks (GCNs) within the LCL module to propagate multi-layer graph information.
• Introduced word embeddings of pathologies to generalize relationships into classifier scores.

Main Results:

• CheXGCN demonstrated superior performance compared to state-of-the-art baselines on benchmark datasets (ChestX-Ray14 and CheXpert).
• The framework effectively recalibrates multi-label outputs by incorporating learned pathology relationships.
• Validation confirmed the effectiveness of explicitly modeling label co-occurrence for CXR classification.

Conclusions:

• CheXGCN offers a powerful approach to enhance multi-label medical image classification by leveraging graph-structured pathology information.
• The proposed method provides a flexible and effective way to integrate label dependencies into deep learning frameworks.
• This work highlights the potential of GCNs in uncovering intricate relationships for improved clinical decision support in radiology.