INFORM-CT: INtegrating LLMs and VLMs FOR Incidental Findings Management in Abdominal CT

Researchers propose INFORM-CT, an AI framework combining large language models and vision-language models to automate detection and reporting of incidental findings in abdominal CT scans. The system uses a planner-executor approach that outperforms traditional manual inspection and existing pure vision-based models in accuracy and efficiency.

INFORM-CT addresses a critical gap in medical imaging workflows where incidental findings—unexpected abnormalities discovered during routine scans—require careful evaluation and documentation. Radiologists currently spend significant time manually inspecting scans, creating bottlenecks and introducing variability in detection quality. This paper demonstrates how combining LLMs and VLMs through an agentic framework automates this process while adhering to clinical guidelines.

The innovation leverages LLMs' reasoning capabilities to generate diagnostic scripts and VLMs' visual understanding to detect and classify findings, creating a hybrid approach more sophisticated than pure computer vision systems. This builds on broader trends in medical AI where multimodal systems demonstrate superior performance compared to single-modality approaches. The planner-executor architecture allows the system to systematically evaluate multiple organ systems according to established medical protocols.

The healthcare sector increasingly seeks AI solutions that reduce radiologist workload while maintaining diagnostic accuracy. Faster incidental finding detection accelerates clinical decision-making and reduces liability from missed diagnoses. Success on abdominal CT benchmarks suggests scalability to other imaging modalities and organ systems, potentially creating market demand for integrated diagnostic AI platforms.

Future applications could extend this framework to real-world clinical deployment, requiring validation across diverse patient populations and integration with existing hospital IT infrastructure. Regulatory approval pathways for clinical use remain a significant hurdle. The technical results indicate practical feasibility, positioning this approach as a meaningful contribution to AI-assisted radiology.

• →INFORM-CT combines LLMs and VLMs in an agentic framework that automates incidental findings detection in CT scans with higher accuracy than existing pure vision-based approaches.
• →The planner-executor architecture enables systematic evaluation of multiple organs according to established clinical guidelines, reducing manual radiologist workload.
• →Multimodal AI systems outperform single-modality approaches in medical imaging tasks, validating the integration of language reasoning with visual understanding.
• →Automated incidental finding detection has potential clinical and economic impact by accelerating diagnosis, reducing liability, and increasing radiologist efficiency.
• →Scaling this technology to clinical practice requires regulatory approval and integration with hospital infrastructure beyond current research benchmarking.