The Model Zoo

This directory is generated from data/models.csv. It includes task-specific models, reusable encoders, and development frameworks. These are different kinds of resources. A framework can help train a model without providing a suitable pretrained checkpoint; a foundation checkpoint can provide features without a validated clinical decision rule.

Modality	Model / framework	Task	Weights	Terms	Scope	Source
cxr	CheXagent 2 (3B)	CXR vision-language interpretation	Open	MIT (checkpoint metadata); research-only stated use	Stanford AIMI; use checkpoint-specific tasks and interface	Source
cxr	TorchXRayVision	Classification (multi-label)	Open	Apache-2.0 code; inspect selected weights and data terms	Research baseline; preserve documented intensity normalization	Source
cxr	MedGemma (medical multimodal family)	VLM (general medical)	Open (gated)	Gemma license	Development starting point; pin version; application-specific validation required	Source
ct	TotalSegmentator	Anatomical segmentation	Default task weights available	Apache-2.0 default model; other tasks may have separate terms	Default CT task has 117 structures; inspect task-specific input and license	Source
mri	fastMRI baselines	Reconstruction	Reference implementations and release-specific checkpoints	MIT code; dataset agreement separate	Research reconstruction baseline; pin sampling and coil setup	Source
us	EchoNet-Dynamic	Video function estimation and segmentation	Reference code and documented model resources	Inspect repository and data/checkpoint terms separately	Apical four-chamber research; not complete echo interpretation	Source
mammo	Mirai	Future breast-cancer risk	Released research model	MIT; authors specify research rather than clinical use	Future-risk task; original code requires supported four-view presentation images	Source
ophtho	RETFound	Retinal representation learning	Checkpoint-specific access	Checkpoint-specific terms; inspect model card	Adapt and validate a task head; not an autonomous screening device	Source
video	TeCNO	Surgical phase recognition	Reference implementation and linked resources	Repository and checkpoint terms apply	Causal temporal research model; phase labels do not assess surgical safety	Source
path	UNI family	Pathology representation learning	Gated checkpoints	Checkpoint-specific noncommercial/research terms	Pin model generation; patch representations need downstream task validation	Source
path	Prov-GigaPath	Pathology representation learning	Gated model resources	Research terms; inspect selected artifacts	Tile and slide representations; validate tissue and stain scope	Source
any	nnU-Net	Self-configuring segmentation	Framework; train or obtain task-specific weights	Apache-2.0 code; task weights and data terms separate	Framework availability does not establish a validated modality checkpoint	Source
any	nnInteractive	Interactive segmentation	Released model resources	Inspect pinned code and model license	Prompt and input-domain validation required	Source
any	BiomedParse	Prompted biomedical segmentation	Released model resources	Code and checkpoint terms differ; inspect model release	Only use supported representations; not universal modality validation	Source
any	MedSAM2	Interactive segmentation	Released model resources	Inspect pinned code and checkpoint terms	Task-specific prompts and evaluation required	Source

A model is more than its weights

A reproducible model package includes architecture, checkpoint checksum, input selection, preprocessing, label map, postprocessing, thresholds, software dependencies, and the inverse mapping to source coordinates. A change in normalization can alter results even when the checkpoint is unchanged.

Treat code, weights, and data licenses separately. Gated access is not itself a license category, and downloadable weights do not imply permission for every commercial or clinical use. Check the exact model card and artifact version. The any category means broadly reusable development resources, not universal validated input support.

Compare models under one task contract

Fix the dataset split, input availability, reference labels, operating point, and evaluation unit before comparing architectures. Include a strong simple baseline. Measure errors that matter to the intended workflow: lesion sensitivity, false-positive review burden, measurement agreement, technical failures, latency, and correction time as applicable.

Record hardware and peak resource use on representative inputs. A model that fits one cropped volume may not fit a full examination with concurrent sessions. Include download and preprocessing in operational timing when they contribute to the user’s wait.

A compact model card for local use

Field	Required content
Identity	Name, version, source, checkpoint checksum
Intended use	Population, modality, task, user, clinical role
Inputs	Acquisition requirements, units, geometry, missing-input policy
Outputs	Labels, scores, calibration, coordinate system, failure states
Evidence	Dataset releases, split policy, metrics, intervals, external evaluation
Limitations	Unsupported cases, known failures, abstention and escalation
Operations	Dependencies, hardware, concurrency, monitoring, rollback
Rights	Code, weights, data, redistribution, and use restrictions

A model-zoo entry is a starting point for investigation. It is not a clinical endorsement or a substitute for validation of the deployed acquisition-to-result chain.