Researchers built a two-stage AI model that reads lung CT scans, describes nodules, and writes follow-up recommendations. That work usually falls to a radiologist.
The system, called FZ-VLM, splits the job in two. A fine-tuned Florence-2 model first extracts attributes like nodule location, margin, and attenuation type from CT slices, then estimates diameter. A second model, Zephyr-7B, takes those attributes and generates a written description, a follow-up recommendation, and a comparison to prior scans. In testing, the extraction stage hit 77.18% accuracy on location, 67.96% on margin characteristics, and 79.13% on attenuation type, with a diameter error of 2.58mm, outperforming GPT-4-based baselines and a human comparison group.
Lung cancer screening produces a steady stream of nodules that need consistent, structured writeups, and inter-observer variability among radiologists is a known problem. A tool that reliably drafts that first pass could cut review time without replacing clinical judgment. Radiologists rated the generated reports 93.9% accurate and 98.6% complete, though clinical relevance lagged at 76.1%.
The margin-characterization accuracy, still under 70%, is a reminder that pattern recognition on CT slices remains hard, and the researchers themselves say some follow-up recommendations still need expert review before they reach a patient.