New AI Framework Automates Detection of Overlooked Findings in Abdominal CT Scans

Incidental findings—unexpected abnormalities discovered during medical imaging—often go unnoticed or are inconsistently reported by radiologists reviewing CT scans. Researchers have developed INFORM-CT, a framework that pairs large language models with vision models to automate the detection, sorting, and reporting of these discoveries in abdominal imaging.

The Challenge with Manual Review#

Identifying incidental findings in abdominal CT scans has traditionally relied on radiologists manually examining images. This approach is labour-intensive, prone to variability between clinicians, and can miss subtle anomalies. Yet these findings, though usually benign, carry important clinical weight and should be documented according to medical standards. The volume of scans and complexity of assessment creates a genuine bottleneck in radiology workflows.

How the System Works#

The INFORM-CT approach uses a two-part "plan-and-execute" strategy. An LLM-based planner reads medical guidelines for abdominal organs and writes Python scripts that outline the checking and detection process. A separate executor then carries out those scripts using vision-language models, segmentation algorithms, and image processing tools to systematically analyze the scan and flag relevant findings.

Rather than relying on a single visual AI model to interpret an entire image, the framework breaks the task into structured steps, letting the language model choreograph which analyses to run and how to synthesize results.

Early Results#

Testing on abdominal CT benchmarks covering three organs showed the framework outperformed standalone vision-model approaches in both accuracy and speed. The system operates in a fully automatic, end-to-end manner without human intervention during the detection phase.

Why It Matters#

Automating incidental finding detection could lighten the radiologist's workload, reduce missed diagnoses, and standardize reporting across clinics. As CT imaging continues to grow in volume, tools that enhance consistency and efficiency in reporting become increasingly valuable.