An open model explains three-dimensional CT scans
On 23 September 2026 NVIDIA introduced NV-Reason-CT, an open vision-language model that reads a CT volume as a three-dimensional input and drafts structured findings plus a written chain of reasoning for chest and abdominal scans. On the public CT-RATE benchmark the authors report a macro-F1 of 0.614 and a macro-AUROC of 0.871, above the published baselines they list. NIH radiologists reviewed reports and traces for plausibility. The announcement states that the model is a research foundation, not an autonomous diagnostic system and not a cleared clinical product.
The record ends here. Everything below this line is invented.
What if a scan came with a reasoning you were expected to audit?
Part one
The Trace in the Volume
Nadine read lungs by scrolling. A study was a stack, and the mass or the collapse or the nothing lived in the relationship between slices, not in any one of them. The new pane beside the stack wrote the scroll out in sentences, region by region, and it was not allowed to sign anything. She was.
The pane was a research tool. The hospital's rule was that she could borrow its sentences if she could point at the slice that earned each one.
Tonight's trace said a large process in the right lower lobe, with volume loss, and bulky nodes. She found the nodes. She did not find the process. Slice after slice, the lobe looked tired and not mass-like.
She went back to the sentence. It had a confidence and a path through the volume, naming slices she had passed.
Nadine left the finding unsigned and marked the trace as the thing the morning reader would have to answer.
What’s real
- NV-Reason-CT reads CT as a volume and writes a reasoning trace
- It led the published CT-RATE baselines the authors list
- NIH radiologists reviewed traces for plausibility
- The model is not a cleared clinical product
What’s invented
- Nadine's rule that every sentence needs a slice
- A trace she cannot match to the lobe
- An unsigned finding left for morning
- A hospital that lets her borrow the sentences
Part two: The Slice She Could Not Find
Part two isn’t on sale yet. Check back soon.
Filed under medicine, medical imaging, radiology, vision language models.