Introducing NV-Reason-CT Open 3D CT VLM for Radiologist Chain-of-Thought Reasoning

NVIDIA introduced NV-Reason-CT, an open research vision language model built for 3D computed tomography analysis. The company said it extends the chain-of-thought reasoning methodology of NV-Reason-CXR, which it said was validated in a multireader clinical study accepted at RSNA 2026, to full volumetric CT. NVIDIA describes NV-Reason-CT as an open research and development foundation, not an autonomous diagnostic system or a cleared clinical product.

The model combines a full 3D vision transformer encoder with a Qwen3.5-4B language model, trained end-to-end on CT data with structured reports, reasoning traces, and multistep visual question answering. NVIDIA said the encoder is adapted from Primus 3D ViT and initialized with Colipri weights, processing CT volumes resampled to 192-cubed voxels at 2 mm isotropic resolution with 8x8x8 patch tokens, producing 13,824 vision tokens passed to the language model with 3D grid coordinates through 3D MRoPE.

Training followed a two-stage pipeline: supervised fine-tuning on roughly 550,000 structured QA examples across chest and abdominal regions, using CT-RATE, NIH CT datasets, and CancerVerse, plus synthetic reasoning data distilled from large language models; then reinforcement learning with Group Relative Policy Optimization using an anatomy-aware reward based on identified abnormality and diagnosis accuracy.

NVIDIA reported state-of-the-art results on the CT-RATE benchmark, with a Macro-F1 of 0.614 and Macro-AUROC of 0.871, saying the model outperformed published 3D contrastive and fused 2D/3D baselines. It said NIH radiologists validated the clinical plausibility of the structured reports and reasoning traces. Baris Turkbey, M.D., F.S.A.R., Senior Clinician at the National Institutes of Health, said the model provides systematic, step-by-step reasoning reflecting how CT studies are reviewed, and that reviewing its thought process rather than only its conclusions makes it possible to trust and act on findings. The model is available via Hugging Face, GitHub, and a web demo.