Knowledge-Guided and Reinforced Selective State Space Model for radiology report generation

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Generation

Background:

• Current AI models for radiology reports often act as "black boxes" with basic reward functions.
• This limits their clinical accuracy and the linguistic quality of generated reports.

Purpose of the Study:

• To improve the clinical accuracy and linguistic quality of automated radiology report generation.
• To address limitations of "black box" models and simplistic reward functions.

Main Methods:

• Proposed a novel Knowledge-Guided and Reinforced Selective State Space Model (KGR-SSM).
• Integrated a medical knowledge graph for semantic understanding of visual features.
• Utilized a Mamba-based encoder for high-resolution image processing.
• Implemented a hybrid reward function optimizing NLG metrics and clinical accuracy.

Main Results:

• KGR-SSM achieved state-of-the-art performance on IU X-ray and MIMIC-CXR datasets.
• Demonstrated significant outperformance over existing methods in linguistic and clinical efficacy.
• Validated the model's effectiveness across comprehensive evaluation metrics.

Conclusions:

• Integrating structured medical knowledge and a clinically-aligned hybrid reward function enhances report accuracy and reliability.
• The KGR-SSM framework offers a robust solution for automated radiology report generation.
• Successfully bridged the gap between technical performance and clinical utility in AI-driven medical reporting.