🩻 AI that reads a CT scan in 3D—and explains its findings

NVIDIA, the NIH’s National Cancer Institute, and the University of Zurich have released NV-Reason-CT, an open model that analyzes full 3D CT volumes and generates reports with explanations.

⚙️ How it works
The model pairs the Qwen3.5-4B language model with Primus, a 3D visual encoder. Each scan becomes 13,824 visual tokens, passed to the language model without further compression. Three-dimensional positional encoding preserves spatial information, helping it distinguish, for example, a finding in the right kidney from one in the left.

📚 How it was trained
Training used 550,000 examples from 70,111 CT volumes. Supervised fine-tuning on radiologists’ analyses was followed by reinforcement learning, with rewards for correctly identifying abnormalities and following the required report structure.

📊 What the results show
On CT-RATE, NV-Reason-CT achieved an average precision of 0.614 across 18 abnormality categories, compared with 0.581 for VoxelFM and **0.398 for CT-CLIP**—without a separate classification head.

In a pilot study with radiologists, scan review and reporting time fell from 26.25 to 13.13 minutes: roughly half.

🧩 Part of a broader medical AI toolkit
NVIDIA’s open medical model family also includes:
• NV-Generate-CTMR — generates synthetic 3D CT and MRI volumes.
• NV-Segment-CTMR — segments organs and lesions.
• NV-Reason-CXR — analyzes chest X-rays.
• NV-Reason-CT — analyzes full 3D CT scans.

🔓 Weights and code are available under OpenMDW-1.1, alongside fine-tuning and reinforcement-learning examples and a web demo.

Promising early results for AI-assisted radiology—with the time savings demonstrated so far in a pilot study.

@science