RADAR: A Difference-Aware Retrieval with Organ-Level Alignment for Change Radiology Report Generation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Radiology

Background:

• Current automated radiology report generation is limited to static, single-image analysis.
• There is a clinical need for longitudinal comparison to monitor disease progression and treatment efficacy.
• Automated analysis of interval changes between radiological images is lacking.

Purpose of the Study:

• Introduce the task of Change Radiology Report Generation (CRRG) for automated comparative analysis.
• Propose RADAR, a novel framework for generating comparative radiology reports describing interval changes.
• Develop a benchmark dataset and evaluation metrics for the CRRG task.

Main Methods:

• RADAR framework integrates deep visual comparison with knowledge-rich text generation.
• Employs an "align first, then compare" strategy with organ-level alignment for robust visual analysis.
• Text generation incorporates soft prompts for visual evidence, a four-step clinical reasoning workflow, and a Knowledge-Infused Generation (KIG) component for factual accuracy.

Main Results:

• RADAR outperforms existing methods on most evaluation metrics in extensive experiments.
• The proposed framework successfully generates interpretable and clinically relevant change reports.
• The introduced benchmark facilitates robust evaluation of CRRG methods.

Conclusions:

• The RADAR framework advances automated radiological assessment from static analysis to dynamic monitoring.
• CRRG enables automated generation of comparative radiology reports for longitudinal patient assessment.
• This work addresses the critical clinical need for automated longitudinal comparison in radiology.