Nvidia introduced NV-Reason-CT Open on 23 September 2026, an open 3D CT vision-language model that generates radiologist-style chain-of-thought reasoning across volumetric chest and abdominal scans, according to a post on the Nvidia developer blog.
Key points
- Nvidia introduced NV-Reason-CT Open on 23 September 2026 as an open 3D CT vision-language model
- The model combines a Qwen3.5-4B LLM with a 3D ViT encoder adapted from Primus and initialised with Colipri weights
- CT volumes are resampled to 192³ voxels at 2 mm isotropic resolution, producing 13,824 vision tokens with 3D grid coordinates
- Training uses a two-stage pipeline: supervised fine-tuning on ~550,000 structured QA examples, then GRPO reinforcement learning with anatomy-aware rewards
- Nvidia states NV-Reason-CT is an open research foundation, not an autonomous diagnostic system or cleared clinical product
Nvidia’s NV-Reason-CT addresses 3D CT volumetric complexity
A single abdominal CT study can contain 300 to 600 axial slices, carrying anatomical information across three spatial planes that a standard 2D encoder cannot rebuild from separate frames. The Nvidia blog post explains that treating a volume as a stack of 2D frames loses the spatial ties between slices that define masses, effusions and infiltrates — structures that only reveal their full nature when viewed volumetrically. Even models that perceive an abnormality often output a diagnostic label without articulating why, and most lack the multiturn dialogue capability to support the iterative clinical reasoning a radiologist performs when reviewing a suspicious finding.
Architecture built for spatial reasoning
NV-Reason-CT combines a Qwen3.5-4B language model with a 3D vision transformer encoder adapted from Primus and initialised with Colipri weights. CT volumes are resampled to 192³ voxels at 2 mm isotropic resolution and divided into non-overlapping 8×8×8 patch tokens, yielding a 24×24×24 grid of 13,824 vision tokens. Unlike approaches that merge or downsize tokens, all tokens are passed to the LLM together with their 3D grid coordinates, and the language model incorporates 3D multiscale rotary positional embeddings (MRoPE) to account for spatial inter-token relationships throughout its layers. This design maintains anatomical continuity across slices and lets the model work with structures as whole entities, much as a radiologist does when moving through a study. The model generates structured reports mapped to a curated ontology of 30 chest and 29 abdominal abnormalities, produces step-by-step reasoning that mirrors a radiologist’s systematic review, and supports multiturn conversational follow-up on specific findings or differential diagnoses.
The model generates step-by-step reasoning that systematically reviews anatomical regions, notes normal and abnormal findings, and expresses calibrated uncertainty, rather than outputting only a final label.
Two-stage training with anatomy-aware reinforcement
The training pipeline follows the methodology pioneered with NV-Reason-CXR. Stage one uses supervised fine-tuning on roughly 550,000 structured question-answer examples from CT-RATE, NIH CT datasets and CancerVerse, augmented with synthetic reasoning data produced from large language models guided by radiologist-reviewed cases. The curriculum spans section-level anatomy QA, laterality-specific and localised finding QA, severity-level QA and binary abnormality identification, with refusal examples for invalid prompts and mismatched image-text pairs. Stage two applies Group Relative Policy Optimization with an anatomy-aware reward function that reinforces accuracy within each anatomical region rather than a single global signal, improving calibration across the full chest-abdomen findings distribution. This two-stage approach — learning reasoning patterns first, then reinforcing correctness — lets the model handle the wide variety of CT cases without needing fully labelled reasoning traces for every possible presentation.
Benchmark performance and clinical positioning
NV-Reason-CT achieves state-of-the-art results on the leading public benchmarks for 3D CT understanding, the Nvidia blog post states. The reasoning methodology builds on NV-Reason-CXR, validated in a multireader clinical study accepted at RSNA 2026 that showed radiologist time savings with diagnostic accuracy preserved. Nvidia emphasises that NV-Reason-CT is an open research foundation, not an autonomous diagnostic system or a cleared clinical product, and is designed for researchers and developers building specialised CT analysis applications to post-train for their use case.
