Skip to main content
Back

NVIDIA's NV-Reason-CT Reads Full 3D Body Scans to Write Radiology Reports

NVIDIA's open 4.69B parameter 3D vision-language model turns full chest and abdominal CT volumes into structured radiology reports in under 30 seconds on a single GPU.

  • NVIDIA released NV-Reason-CT, an open 4.69B 3D vision-language model for chest and abdominal CT.
  • Native 3D ViT (Primus/COLIPRI) plus Qwen3.5-4B, with 13,824 spatial tokens preserved via 3D MRoPE.
  • Trained with SFT and GRPO on ~550K QA examples from 70,111 CT volumes.
  • Beats VoxelFM, Pillar-0, and MedGemma 1.5 on CT-RATE classification (0.614 Macro-F1).
  • Runs on a single H100, L40S, or A100 under the permissive OpenMDW-1.1 license.
  • Research and education only; code and paper available.

NVIDIA has released NV-Reason-CT, an open-weights vision-language model that processes full 3D CT volumes. The checkpoint can generate structured chest and abdominal radiology reports, answer follow-up questions, and produce reviewable reasoning alongside its answers.

Parameters
4.69 billion total, with 4.35 billion in the active CT pathway
Inputs
Chest or abdominal CT volumes in NIfTI format with valid Hounsfield-unit values
Outputs
Structured reports, abnormality questions, multi-turn answers, and generated reasoning
Deployment
Single-GPU inference tested on NVIDIA H100, L40S, and A100 hardware
License
OpenMDW 1.1

Why full-volume context survives

Many medical vision-language systems process selected 2D slices or compress volumetric features before sending them to a language model. Those methods can lose spatial relationships across slices, including the extent, distribution, and relative position of findings.

NV-Reason-CT uses a native 3D vision transformer called Primus, initialized with COLIPRI, alongside Qwen3.5-4B. The encoder converts a 384 × 384 × 384 mm region into a 24 × 24 × 24 grid containing 13,824 visual tokens. Every token reaches the language model with explicit depth, height, and width coordinates encoded through a 3D form of rotary positional embedding called MRoPE.

Input volumes are cropped to the requested chest or abdominal region with a lung Hounsfield-unit heuristic, then resampled to 2 mm isotropic resolution. Isotropic resampling gives each voxel the same physical dimensions along all three axes, which simplifies 3D processing and preserves consistent spatial scale.

How expert readings shape training

Training combines supervised fine-tuning with Group Relative Policy Optimization, a reinforcement-learning method that compares candidate responses using defined rewards. NVIDIA reports approximately 550,000 structured question-and-answer examples derived from 70,111 unique CT volume inputs.