License-preflight scaffold for a planned commercial-use-friendly brain-MR multimodal model. No weights, training code, or serving code exist yet.
Maintenance status (2026-09): passive. This repository is kept available as a reference implementation; CI runs on pushes and pull requests only, Dependabot security alerts remain enabled, and no scheduled jobs or hosted services consume ongoing resources. No active development is planned.
medimage-model is the fully-permissive track of the GOATnote medical imaging stack. Its purpose is to train only on commercial-use-friendly data (OpenNeuro CC0, permissive open corpora, NVIDIA Open Model License synthesis), enforced by a license preflight gate at ingest. Today, the gate and two seed manifests are what exists.
Status: pre-v0.1 scaffold. There is no model here yet — no checkpoints, no training or serving entry points, no eval adapter. The research-track sibling is
medimage-model-research— same scaffold, different data-license boundary.
Nothing in this repository — and no model eventually published from it — is a medical device or a substitute for clinical judgment. Outputs of any future model must not be used for diagnosis or treatment decisions without a separate, regulated clearance process. See medimage_model.CLINICAL_DISCLAIMER, which any future serving path is required to emit.
- License preflight gate (
scripts/license_preflight.py) — SPDX allowlist + forbidden-substring blocklist +commercial_useflag check overdata/permissive_only/manifests/; rejects CC BY-NC-, CC BY-ND-, research-only, and PhysioNet-credentialed terms with exit 2; fails closed when the manifest directory is missing or empty. Runs in CI on every push and PR, with a negative NC fixture exercised end-to-end. - Manifest schema (
src/medimage_model/data/manifest.py) — pydantic v2DatasetManifest/DatasetLicensemirror of the gate's contract. - Seed manifests — OpenNeuro
ds002785(CC0, verified against OpenNeuro metadata) and UCSF-BMSR (recorded as CC BY 4.0; verify on the UCSF portal before this becomes a training input).
The intended stack, none of which exists in this repository today:
- Image encoder: MAISI-v2 latents; MedImageInsight / MedSAM2 backbones for other modalities
- Text decoder: Nemotron-3-Nano-30B-A3B via PEFT/LoRA
- Synthesis: NV-Generate-MR-Brain for long-tail augmentation with strata caps and per-sample tagging
- Training: Megatron + SGLang + GRPO with verifiable rewards
- Eval:
medimage-evalsubstrate (declared as a pinned optional extra; not yet consumed by any code here) - Receipted releases: per-release attestation of code commit, data manifest hash, eval results hash
Most large open medical-imaging paired-report corpora (MR-RATE, CT-RATE) are CC BY-NC-SA. Models trained on them carry a research-only cloud. This track avoids that: every dataset entering training must pass the preflight gate first. See docs/LICENSE_MAP.md.
Not on PyPI (and there is little to install yet — the package ships the manifest schema and a CLI stub):
pip install git+https://github.com/GOATnote-Inc/medimage-model.git@mainRun the gate:
python scripts/license_preflight.py data/permissive_only/manifestsCode: Apache License 2.0 (full text in LICENSE; scope notes in NOTICE). Any future trained weights would be released under their own documented license per release. Training-data licenses: see docs/LICENSE_MAP.md.