Anatomy-Guided Radiology Report Generation With Pathology-Aware Regional Prompts

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Natural Language Processing

Background:

• Radiology report generation aims to reduce workload but faces challenges with complex images and subtle pathologies.
• Current methods struggle to achieve high clinical accuracy due to intricate anatomical structures and pathologies.
• Accurate and efficient radiology reporting is crucial for effective patient management.

Purpose of the Study:

• To develop an AI model for automated radiology report generation with enhanced clinical accuracy.
• To integrate anatomical and pathological information explicitly into the report generation process.
• To improve the linguistic fluency and medical accuracy of generated radiology reports.

Main Methods:

• Developed an anatomical region detector to extract structured visual features from specific areas.
• Implemented a multi-label pathology detector to identify global abnormalities.
• Utilized pathology-aware regional prompts to integrate anatomical and pathological insights for report decoding.

Main Results:

• The model achieved superior performance in natural language generation and clinical efficacy.
• Achieved BLEU-1 score of 0.394, ROUGE-L of 0.302, and F1 score of 0.470.
• Expert evaluations confirmed the model's potential to enhance radiology practice and clinical decision-making.

Conclusions:

• Integrating anatomical and pathological insights emulates radiologists' workflow for improved reporting.
• The model demonstrates superior accuracy and clinical coherence in radiology report generation.
• This approach shows significant promise for supporting clinical decisions and transforming patient management.