NVIDIA's NV-Reason-CT reads 3D CT scans and shows its work, radiologist style

NVIDIA releases NV-Reason-CT, an open model that reads CT volumes and shows its work

NVIDIA published a technical blog post on September 23, 2026 introducing NV-Reason-CT, an open vision-language model built to read full three-dimensional CT scans and generate reasoning written in the style of a working radiologist. The post was authored by NVIDIA researchers Andriy Myronenko, Supriya V. Thathachary and Monty Zarrouk.

The model generates structured diagnostic reports, produces step-by-step chain-of-thought reasoning that walks through anatomical regions one by one, and supports multistep follow-up conversation across chest and abdominal scans. NVIDIA describes it as an open research and development foundation model for researchers and developers, explicitly stating that it is not an autonomous diagnostic system or a cleared clinical product.

The work is a direct extension of NV-Reason-CXR, NVIDIA's chest X-ray reasoning model released in late 2025, whose reasoning methodology was validated in a multireader clinical study that NVIDIA says has been accepted at RSNA 2026.

Why CT is hard for AI: hundreds of stacked images, not one flat photo

A CT scan is not one image. A single abdominal CT study can contain 300 to 600 axial slices, thin cross-sections stacked along the body's length. Radiologists interpret this as a continuous three-dimensional volume: a nodule's shape, an effusion's extent, or a mass's density only become clinically meaningful when the slices are read together.

Most medical AI does not work that way. Standard vision-language models process images as a flat 2D grid of tokens, and even models adapted for CT often treat each slice independently, discarding the spatial relationships between them. General-purpose frontier models, NVIDIA reports, perform poorly on volumetric imaging.

This matters beyond perception. Even models that correctly spot an abnormality frequently output a diagnostic label without explaining why. Radiologists, by contrast, think in systematic anatomical reviews, differential diagnoses, and calibrated degrees of confidence. A model that cannot reproduce that reasoning process cannot be audited, taught from, or safely integrated into clinical workflows.

A purpose-built 3D architecture, not a 2D model with slices glued together

The technical core of NV-Reason-CT is that it processes a CT volume as genuinely three-dimensional input rather than as a stack of independent 2D frames. The encoder is a full 3D vision transformer adapted from NVIDIA's Primus architecture and initialized with Colipri weights. CT volumes are resampled to 192 cubic voxels at 2 mm isotropic resolution and divided into non-overlapping 8x8x8 patch tokens, producing a context of 24 by 24 by 24, or 13,824, vision tokens.

Unlike many designs that merge or downsize visual tokens before the language model sees them, all 13,824 tokens are passed directly to the language model, together with their 3D grid coordinates. The Qwen3.5-4B language model uses 3D MRoPE, a rotary position encoding that accounts for three-dimensional spatial relationships among the vision tokens throughout its layers. All weights are retrained end-to-end on large cohorts of CT data with structured reports, reasoning traces, and internally designed multistep visual question answering.

In practical terms, the model can reason about spatial extent, cross-sectional morphology, and relationships between slices, which NVIDIA describes as the perceptual foundations of accurate CT interpretation.

How the model learned to think out loud

Training followed the two-stage pipeline pioneered with NV-Reason-CXR: supervised fine-tuning followed by reinforcement learning.

In stage one, the model learned from a mixture of structured reports, expert radiologist reasoning annotations, and general visual question answering. Radiologists contributed detailed chain-of-thought dictations that capture their internal review process: what they examine in each anatomical region, which findings they consider significant, which differential diagnoses they weigh, and how they reach a final assessment. The curriculum spans approximately 550,000 structured QA examples across chest and abdominal regions, covering section-level anatomy questions, laterality-specific and localized finding questions, severity-level questions, and binary abnormality identification. Training data includes CT-RATE, NIH CT datasets, and CancerVerse, supplemented with synthetic reasoning data distilled from large language models and grounded in the expert annotations.

In stage two, the team used Group Relative Policy Optimization (GRPO), a reinforcement learning method, to refine reasoning quality. The reward is anatomy-aware: the model is reinforced for accuracy within each anatomical region rather than on a single global score, which NVIDIA says improves calibration across the full chest and abdomen findings distribution. The design intent, per the blog post, is a model that responds not as a classifier but as a teacher, explaining the problem, walking through the evidence, and arriving at a conclusion through visible logical steps.

Benchmark results: reported state of the art on CT-RATE

To guide and evaluate the model, the NVIDIA team curated a CT ontology covering 30 chest and 29 abdominal abnormalities, including lung nodules, pneumothorax, hepatic lesions, and renal cysts, in a format that maps to clinical documentation workflows.

On CT-RATE, the leading public benchmark for 3D CT understanding, NVIDIA reports that NV-Reason-CT outperforms all published baselines, including 3D contrastive models (VoxelFM, Pillar-0, CT-CLIP, Merlin), fused 2D/3D multimodal models (ClinFusion-8B), and slice-based frontier models (MedGemma 1.5). On the benchmark's 18-label classification task with a fixed uniform threshold, NV-Reason-CT scored a Macro-F1 of 0.614 and a Macro-AUROC of 0.871. VoxelFM scored 0.581 and 0.870; Pillar-0 scored 0.544 and 0.861; ClinFusion-8B scored 0.442; CT-CLIP scored 0.398 and 0.733; Merlin scored 0.358 and 0.662; MedGemma 1.5 scored 0.303. NVIDIA notes this is the first time a single open model has achieved competitive CT classification and report generation simultaneously, and that the model was evaluated with a direct Yes/No prompt without a classification head or task-specific adaptation.

These numbers come from NVIDIA's own evaluation and should be read as such. They are strong relative results on a public benchmark, not proof of clinical performance.

What NIH radiologists said

Beyond the benchmark, radiologists at the National Institutes of Health reviewed the model's structured reports and reasoning traces and validated their clinical plausibility. The blog post quotes Baris Turkbey, M.D., F.S.A.R., Senior Clinician at the National Institutes of Health:

"NV-Reason-CT provides the kind of systematic, step-by-step reasoning that reflects how we actually think through a CT study. Being able to review the model's thought process, not just its conclusions, is what makes it possible to trust and act on its findings."

Turkbey statement, September 23, 2026, NVIDIA Technical Blog, "Introducing NV-Reason-CT: Open 3D CT VLM for Radiologist Chain-of-Thought Reasoning". This is an expert review of output quality, not a clinical trial: it indicates that experienced radiologists found the reasoning plausible, not that the model improves patient outcomes. In the earlier NV-Reason-CXR work, Turkbey also highlighted the model's potential as an educational assistant for radiology and medical trainees, context that suggests where clinicians see this line of research heading.

NVIDIA also highlights time savings in structured reporting as a reviewed benefit, noting that a detailed structured report covering more than 60 abnormalities is time-consuming even for experienced radiologists.

What is observed, what is predicted, what is opinion

The observed capabilities are concrete: a released open model that natively ingests 3D CT volumes, produces structured reports and radiologist-style reasoning, handles follow-up questions, and posts the best published scores on CT-RATE with positive expert review of its outputs.

Predictions and opinion, labeled as such

What is prediction is everything downstream of that. NVIDIA positions NV-Reason-CT as a foundation for post-training toward specialized CT applications, and its claim that radiologists saved time with the predecessor chest X-ray model is based on a study accepted at RSNA 2026, not on deployment evidence for this CT model. Whether reasoning-trace auditability translates into safer clinical AI in practice, whether benchmark gains hold in hospital populations, and whether regulatory pathways accept chain-of-thought outputs are all open questions. No cleared diagnostic product exists here today.

Our opinion, clearly labeled as such: the most consequential idea in this release is not the benchmark number but the auditability. An AI whose reasoning a doctor can read, check, and disagree with is a fundamentally different kind of medical AI than a black-box classifier, and it is the right direction for a field where trust has been the bottleneck. But plausibility review by a handful of NIH radiologists is a beginning, not an endpoint, and nothing about patient care changes until independent clinical validation and regulatory review say otherwise.