NV-Reason-CT: 3D Visual Language Model for CT Analysis

The model was trained on real radiologist expertise: NV-Reason-CT was developed using approximately 550,000 training examples derived from actual radiologist interpretations, including recorded expert reasoning and transcribed CT analyses. This means the AI learned from how experienced doctors actually think through medical imaging, rather than just pattern-matching to reports.

Transparency in medical reasoning is built in: Unlike many AI systems that produce a diagnosis as a "black box," this model supports "reviewable observations, differential diagnoses, and uncertainty"—meaning it can show its reasoning steps and acknowledge what it's unsure about, which is important for doctors to maintain oversight and trust.

The model preserves 3D spatial information: By using native 3D visual encoding and passing explicit 3D coordinates through the entire system, NV-Reason-CT retains the full volumetric structure of CT scans rather than flattening them into 2D slices. This technical choice helps the model understand spatial relationships that matter for accurate medical diagnosis.

Performance was validated on independent datasets: The model was tested on both public CT benchmarks and a separate held-out NIH cohort, achieving a macro-F1 of 0.614 and macro-AUROC of 0.871 on CT-RATE. Testing on data the model hadn't seen during training helps verify that performance isn't inflated and that the system generalizes to real-world cases.

The system supports interactive clinical workflows: NV-Reason-CT enables multi-turn interactions and can generate full reports, suggesting it's designed to integrate into actual radiologist workflows rather than replace human judgment entirely.