Boosting lesion annotation via aggregating explicit relations in external medical knowledge graph.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Graph Neural Networks

Background:

• Accurate multi-label annotation of chest X-rays is challenging due to visual-textual modality gaps.
• Existing methods using Graph Convolutional Networks (GCN) struggle with label imbalance and external medical knowledge integration.

Purpose of the Study:

• To develop a Knowledge Graph (KG)-augmented model for improved multi-label lesion annotation on chest X-ray images.
• To address limitations of current GCN approaches by incorporating explicit label relationships from external KGs.

Main Methods:

• Proposed AER-GCN model, leveraging GCN to learn explicit label relations from a medical KG (SNOMED CT).
• Aggregated explicit KG relations with a statistical graph derived from label co-occurrence matrix.
• Introduced three methods for modeling explicit label correlations and two for incorporating these into co-occurrence relations.

Main Results:

• AER-GCN demonstrated superior performance in multi-label lesion annotation compared to state-of-the-art models.
• Evaluated on ChestX-ray and IU X-ray datasets, showing significant improvements.

Conclusions:

• Jointly exploiting dataset co-occurrence and external KG label relations enhances multi-label annotation accuracy.
• The AER-GCN model offers a promising approach for comprehensive chest X-ray interpretation.