Vision-language models as an integrative layer for clinical artificial intelligence in radiology: A systems-level...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Computer Vision
• Natural Language Processing

Background:

• Current AI in radiology uses isolated, task-specific tools with limited clinical context.
• Vision-language models (VLMs) integrate visual and textual data, offering a potential solution for better context and summarization.

Purpose of the Study:

• To propose VLMs as a semantic layer for existing AI orchestration infrastructure in radiology.
• To critically appraise the capabilities and limitations of current state-of-the-art radiology VLMs.
• To identify technical requirements for effective VLM implementation in clinical settings.

Main Methods:

• Review of existing AI orchestration middleware, reporting standards (DICOM SR, HL7 FHIR), and knowledge graphs.
• Critical appraisal of three leading VLM systems (MAIRA-2, RadioRAG, foundation model) using benchmark data.
• Analysis of VLM performance metrics including logical precision, retrieval gains, latency, and demographic bias.

Main Results:

• VLMs uniquely add capabilities in semantic interpretation and summarization but underperform in areas like logical precision (~52-56%).
• Evaluated VLMs demonstrated significant latency increases and systematic demographic bias.
• Most VLM capabilities are still in research or early-pilot stages, not ready for widespread clinical deployment.

Conclusions:

• VLMs show promise as a coordination layer in radiology AI, bridging isolated systems.
• Significant technical challenges including interoperability, bias mitigation, and latency must be addressed.
• The value of VLMs lies in augmenting, not replacing, existing integration infrastructure.