NVIDIA has released NV-Reason-CT, a research model designed to analyze chest and abdominal computed-tomography (CT) scans as complete 3D volumes. A finding can extend across many neighboring slices; treating those images independently can discard the spatial continuity that helps define its shape and extent.
NV-Reason-CT pairs a 3D vision transformer, which encodes the scan, with Qwen3.5-4B, a language model. NVIDIA says the vision encoder divides a volume into small 3D blocks and passes their spatial coordinates with their image features to the language model. That lets the system use relationships across all three dimensions instead of receiving a flattened series of unrelated 2D images.
The model can produce structured reports and answer follow-up questions about chest and abdominal findings. NVIDIA says it trained the model on about 550,000 structured question-and-answer examples, then used a second stage of reinforcement learning to reward accurate identification of abnormalities and diagnoses.
On CT-RATE’s 18-label chest-abnormality task, NVIDIA reports macro-F1 of 0.614 and macro-AUROC of 0.871. Macro-F1 averages each label’s precision-and-recall balance; macro-AUROC averages how well the model ranks scans with an abnormality above scans without one. The AUROC is close to VoxelFM’s 0.870 in its paper, but the evaluation setups differ: that study used a separately trained classifier for each label, while NVIDIA says NV-Reason-CT used direct yes-or-no prompts, a fixed threshold and no task-specific classifier.
NVIDIA says NIH radiologists reviewed sample reports and reasoning traces for plausibility, but the release does not report a prospective clinical evaluation. Its repository labels NV-Reason-CT as research-and-development software, not a medical device, and says qualified professionals must review its outputs.
Sources
- NVIDIA Technical Blog: NV-Reason-CT architecture, training and benchmark results
- NVIDIA-Medtech/NV-Reason-CT repository and research-use notice
- VoxelFM paper: CT-RATE evaluation method and AUROC
- Lead image: Mikael Häggström, “Abdominal CT with scan range and field of view, without annotations”, CC0 1.0.
