Text-image alignment for ILD imaging: linking CXR evidence to CT quantification.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonology

Background:

• Interstitial Lung Disease (ILD) assessment uses chest radiographs (CXR) for screening and computed tomography (CT) for quantification.
• Current AI models often analyze CXR and CT independently, causing report inaccuracies and inconsistencies.

Purpose of the Study:

• To develop a unified AI framework (ARCTIC-ILD) that integrates CXR-derived textual evidence with CT-based segmentation and quantification for ILD.
• To improve the accuracy and consistency of AI-assisted ILD reporting by bridging the gap between imaging modalities.

Main Methods:

• Developed ARCTIC-ILD, a framework that first extracts structured findings from CXR using a calibrated evidence extractor.
• Utilized a terminology-to-mask module with cross-attention adapters to generate lobe-aware CT masks and burden estimates conditioned by CXR findings.
• Implemented an explicit vision-language audit to ensure consistency between textual reports and quantitative data.

Main Results:

• ARCTIC-ILD significantly reduces text hallucination in AI-generated reports.
• The framework demonstrates improved phrase-to-mask alignment between CXR evidence and CT quantification.
• No additional inference latency was observed compared to existing methods.

Conclusions:

• ARCTIC-ILD effectively aligns CXR-derived textual information with CT quantification for ILD assessment.
• This auditable approach integrates reporting and quantification, supporting clinical workflows for ILD triage, structured reporting, and monitoring.