NV-Reason-CT
An open-source 3D CT vision-language model released by NVIDIA that supports radiologist-style chain-of-thought reasoning. It processes three-dimensional volumetric data and produces interpretable diagnostic reasoning outputs, targeting the medical imaging AI domain.
Timeline (last 90 days)
NV-Reason-CT's model weights are hosted on Hugging Face, the code and fine-tuning scripts along with accompanying skills are published on GitHub, and the paper can be accessed on arXiv.
NV-Reason-CT training consists of two stages: first supervised fine-tuning, followed by reinforcement learning optimization, with the reward signal targeting whether imaging findings are correctly identified.
On the CT-RATE Abnormality benchmark, NV-Reason-CT achieved the highest macro F1 score in the comparison using only simple yes/no prompts and without relying on a classification head.
NV-Reason-CT supports flexible interaction modes: the reasoning process can be turned off for fast yes/no answers, or follow-up questions can be asked continuously about the same scan.
NVIDIA open-sourced a medical imaging model called NV-Reason-CT, which can read CT scans in native 3D and reason step by step to generate radiology reports.
At the core of NV-Reason-CT is a native 3D visual encoder whose output is fed into the QWEN 3.5 language model, with the two components trained jointly end-to-end.
During the data preprocessing stage, a processor called MoniTransforms is used to load scans, resample them, and automatically crop out the chest or abdominal region.
NV-Reason-CT's visual encoder divides the volumetric data into patches on a 3D grid, generating a total of 13,824 visual tokens, each of which is sent into the language model together with its three-dimensional spatial position marker.
NV-Reason-CT's training uses approximately 550,000 question-answer samples derived from approximately 70,000 CT volumes, with the reasoning process guided by radiologists' recorded image interpretation results.
NV-Reason-CT introduces a radiologist-style chain-of-thought reasoning mechanism that simulates a doctor's step-by-step inference process.
2 more timeline events
All Facts (12)
NV-Reason-CT's model weights are hosted on Hugging Face, the code and fine-tuning scripts along with accompanying skills are published on GitHub, and the paper can be accessed on arXiv.
50%UnverifiedNV-Reason-CT's training uses approximately 550,000 question-answer samples derived from approximately 70,000 CT volumes, with the reasoning process guided by radiologists' recorded image interpretation results.
50%UnverifiedNV-Reason-CT training consists of two stages: first supervised fine-tuning, followed by reinforcement learning optimization, with the reward signal targeting whether imaging findings are correctly identified.
50%UnverifiedOn the CT-RATE Abnormality benchmark, NV-Reason-CT achieved the highest macro F1 score in the comparison using only simple yes/no prompts and without relying on a classification head.
50%UnverifiedNV-Reason-CT supports flexible interaction modes: the reasoning process can be turned off for fast yes/no answers, or follow-up questions can be asked continuously about the same scan.
50%UnverifiedAt the core of NV-Reason-CT is a native 3D visual encoder whose output is fed into the QWEN 3.5 language model, with the two components trained jointly end-to-end.
50%UnverifiedDuring the data preprocessing stage, a processor called MoniTransforms is used to load scans, resample them, and automatically crop out the chest or abdominal region.
50%UnverifiedNV-Reason-CT's visual encoder divides the volumetric data into patches on a 3D grid, generating a total of 13,824 visual tokens, each of which is sent into the language model together with its three-dimensional spatial position marker.
50%UnverifiedNV-Reason-CT introduces a radiologist-style chain-of-thought reasoning mechanism that simulates a doctor's step-by-step inference process.
50%UnverifiedAs a newly released model, NV-Reason-CT's real-world clinical performance still needs to be validated on actual data from multiple centers and disease types.
50%UnverifiedNVIDIA open-sourced a medical imaging model called NV-Reason-CT, which can read CT scans in native 3D and reason step by step to generate radiology reports.
50%UnverifiedNVIDIA released NV-Reason-CT, an open-source 3D CT vision-language model that supports chain-of-thought reasoning.
50%