Attribute Prototype-Guided Iterative Scene Graph for Explainable Radiology Report Generation.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology Report Generation
• Computer Vision

Background:

• Automated radiology report generation shows promise in reducing radiologist workload.
• Current methods using anatomical regions improve interpretability but struggle with relation reasoning and shared attributes.

Purpose of the Study:

• To propose a novel framework, Attribute Prototype-guided Iterative Scene Graph generation (AP-ISG), to enhance automated radiology report generation.
• To improve interpretability and relational reasoning by using scene graph generation as an auxiliary task.

Main Methods:

• Developed the AP-ISG framework incorporating an Iterative Scene Graph Generation (ISGG) module and an Attribute Prototype-guided Learning (APL) module.
• ISGG utilizes an autoregressive scheme for structural edge reasoning and contextualization for relational reasoning.
• APL optimizes attribute commonalities among regions by enhancing intra-prototype matching and reducing inter-prototype semantic overlap.

Main Results:

• The AP-ISG framework demonstrated superior performance across multiple metrics in experiments.
• The approach effectively models attribute commonalities among anatomical regions.

Conclusions:

• The proposed AP-ISG framework significantly advances automated radiology report generation.
• Integrating scene graph generation enhances interpretability and relational reasoning capabilities in medical imaging analysis.