MedVTok
General-purpose medical visual tokenizer that encodes 2D images and 3D volumes into one token space (global features, patch tokens and latents) for classification, retrieval, segmentation, synthesis, VQA and report generation; trained on over 28M images and 70K volumes
LICMITUPD2026-10-27STATUSMaintainedCOMMPermissive
- What it is
- Foundation model (Foundation models & agents)
- Modality
- Multimodal: 2D images and 3D CT/MRI volumes (9 modalities incl. X-ray, ultrasound, pathology)
- Maintainer
- Ma, Ji, Ning, Liu et al. (Masaaki-75)
- License
- MIT
- Commercial use
- Permissive
- Access
- Free, public source repository (see links below)
- Last update
- 2026-10-27 (last repository push, UTC; Maintained)
- Last checked
- 2026-11-16
- Limitations
- Listed for research and development reference. Hounsfield Lab has not tested its clinical performance; check the project’s own documentation for intended use and regulatory status.
Links
License review: Repo LICENSE and the Hugging Face model card (massaki75/medvtok, public, not gated, checkpoint downloaded with hf download) state MIT. Training uses features from several expert encoders and public datasets that carry their own terms. Paper: arXiv 2610.35842 (Oct 26, 2026).
Related tools in Foundation models & agents
- Pillar-0 (finetune code + HF weights): Open 3D radiology foundation model; finetuning framework, weights on Hugging Face
- Pillar-0 pretraining: Vision-language pretraining code for Pillar-0
- RATE: Uses an LLM (Qwen3 via sglang) to extract structured labels for 366 findings from radiology reports
- Merlin: 3D vision-language model for CT that uses structured EHR and report supervision
- MedGemma: Google's open medical model family; repo has notebooks and code
- CheXagent: Foundation model for chest X-ray interpretation