NV-Reason-CT — 3D Chest & Abdominal CT Reasoning

Run this model →
Version 0.1 · Licence OpenMDW-1.1 · Released 2026-09-25
CT

NV-Reason-CT reads 3D chest and abdominal CT scans for research. Choose a structured report organised by anatomy, or a step-by-step reasoning walkthrough that links observations to differential diagnoses and states its uncertainty.

Available tasks (4)

Structured chest CT report

Default

A report organised by anatomy (devices, lungs and airways, pleura, mediastinum, heart, vessels, esophagus, chest wall, bones) with a numbered impression, from a crop centred on the lungs.

Input
3D CT scan covering the chest.
Output
Structured report organised by anatomy, with a numbered impression and a findings list (.txt and JSON).

Structured abdominal CT report

A report organised by organ (liver and biliary system, spleen, pancreas, adrenals, kidneys and urinary tract, bowel, peritoneum, lymph nodes, vessels, abdominal wall, bones) with a numbered impression, from a crop placed below the lung bases.

Input
3D CT scan covering the abdomen from the diaphragm down.
Output
Structured report organised by anatomy, with a numbered impression and a findings list (.txt and JSON).

Chest CT reasoning walkthrough

A first-person review of the chest in a radiologist's search pattern, linking positive and negative observations to differential diagnoses and stating uncertainty.

Input
3D CT scan covering the chest.
Output
Step-by-step reasoning walkthrough with a findings list (.txt and JSON).

Abdominal CT reasoning walkthrough

A first-person review of the abdomen in a radiologist's search pattern, linking positive and negative observations to differential diagnoses and stating uncertainty.

Input
3D CT scan covering the abdomen from the diaphragm down.
Output
Step-by-step reasoning walkthrough with a findings list (.txt and JSON).

Research details

Training data
About 550,000 instruction examples built from 70,111 CT scans and their radiology reports. Supervision combines structured reports, finding-specific questions, multi-turn dialogue, transcribed radiologist reading sessions and reasoning rewritten from reports in their style. Supervised fine-tuning was followed by reinforcement learning (GRPO) rewarded on agreement with 30 chest and 29 abdominal finding labels.
Intended use
Research and education on chest and abdominal CT: drafting structured reports and reviewable reasoning for comparison with expert reads. Not for diagnosis, triage or treatment decisions.
Known limitations
Each task reads one fixed 384 mm cube resampled to 2 mm, so findings outside it, and detail finer than the 2 mm grid, are not seen. The cube is placed by finding the lungs, so a scan without enough lung in view (for the abdomen, the lung bases) is refused with an explanation rather than read. Like other generative models it can state findings that are not in the image, and its reasoning text is an explanation for review, not a trace of how the model computed its answer. It does not use clinical history or prior scans. Trained on adult CT from a few institutions; rare findings are under-represented.

Citations & references

Citing Nalvera.AI

Please cite the platform alongside the model's own paper listed above — the model authors deserve the credit for the method.

@misc{nalvera,
  title        = {Nalvera.AI: a cloud platform for medical-imaging AI},
  author       = {{van der Heyden, Brent}},
  year         = {2026},
  howpublished = {\url{https://nalvera.ai}},
  note         = {Model: NV-Reason-CT — 3D Chest & Abdominal CT Reasoning v0.1. Accessed YYYY-MM-DD}
}
Research use only — not a medical device. Outputs are not validated for clinical diagnosis or treatment.
Run this model → ← Back to models