Beyond the Visible
Rethinking Healthcare Vision AI for an Open-Vocabulary World
Adaptive Clinical Visual Intelligence — From Closed-Set Detection to Evidence-First, Agentic and Continuously Learning Healthcare AI
Healthcare is becoming increasingly visual. Radiographs, CT and MRI studies, ultrasound, pathology slides, retinal images, dermatology photographs, endoscopic procedures, surgical video, wound photographs, dental images, smartphone-based screening and continuous bedside video collectively represent an enormous and rapidly expanding source of clinical evidence.
Artificial intelligence has demonstrated substantial potential in extracting information from these sources. Yet much of Healthcare Vision AI continues to operate within a paradigm inherited from conventional supervised machine learning:
Define the clinical problem, create a label taxonomy, collect examples, annotate them, train a model and deploy the model to recognize the same predefined labels. This paradigm remains appropriate for many bounded clinical applications. Its limitation is structural rather than methodological: the model’s operational vocabulary is largely determined before deployment, while clinical reality is not.
Medical vision-language models (VLMs) introduce a fundamentally different capability. By aligning medical images and clinical language within related representation spaces, they enable semantic retrieval and open-vocabulary or zero-shot evaluation. MedSigLIP, for example, is documented as supporting zero-shot classification and semantic medical-image retrieval across a range of medical imaging domains.
Healthcare should not interpret this development simply as a new classification technique. It should use it to reconsider the architecture of clinical visual intelligence as Adaptive Open-Vocabulary Clinical Vision Intelligence — AOV-CVI.
AOV-CVI is a model-agnostic, evidence-first, open-vocabulary, temporal, agentic, safety-controlled and lifecycle-governed paradigm for converting healthcare images and video into computable clinical intelligence.
Its objective is not to replace specialist medical-imaging models with a single foundation model. Nor is it to permit unrestricted zero-shot diagnosis. Instead, AOV-CVI establishes a hybrid intelligence architecture in which specialist Computer Vision, medical VLMs, zero-shot discovery, temporal models, normality models, visual retrieval, clinical knowledge, human expertise and governed learning operate together.
Its core philosophy is: See → Understand → Remember → Relate → Reason → Verify → Act → Learn → Govern
The paradigm introduces five major architectural constructs:
- Clinical Visual Intelligence Fabric (CVIF) — a model-independent intelligence fabric for generating and reasoning over spatial, semantic, morphological, temporal and longitudinal visual evidence.
- Clinical Visual Memory (CVM) — governed longitudinal visual memory spanning patients, anatomical structures, lesions, encounters, procedures, reference populations and outcomes.
- Clinical Visual Semantic Fabric (CVSF) — an ontology, terminology and knowledge-graph layer that transforms visual evidence into computable clinical knowledge.
- Agentic Clinical Vision Control Plane (ACVCP) — specialized agents that orchestrate acquisition, quality, evidence generation, model routing, memory, semantic reasoning, retrieval, safety, review and governance.
- Open-to-Known Clinical Learning Loop (OKCL) — a MedOps-governed mechanism through which previously unmodeled patterns surfaced by open-vocabulary intelligence can be clinically qualified, incorporated into the governed vocabulary, annotated, independently validated and promoted into known production capabilities.
The central proposition is therefore: (a) Zero-shot intelligence discovers. (b) Clinical experts qualify. (c) MedOps industrializes. (d) Validated AI operationalizes. (e) Continuous monitoring provides the evidence for the next learning cycle.
The result is not simply another AI model. It is a learning clinical visual intelligence ecosystem.
1. Healthcare Has a Vision Problem
Healthcare does not suffer from a shortage of visual data. It suffers from fragmentation in how that visual evidence is acquired, represented, interpreted, understood, remembered and governed.
A chest radiograph may be interpreted by one application, a pathology slide by another, a skin lesion photograph by another, while endoscopic and surgical video may follow separate pipelines. Each domain can have its own model, label taxonomy, evidence representation and lifecycle.
Specialization is necessary because modalities differ. A CT volume, whole-slide pathology image and smartphone photograph should not be forced through identical algorithms. But specialization does not require architectural isolation. A common intelligence fabric can sit above modality-specific capabilities and normalize their outputs as evidence.
The opportunity is therefore not a universal medical-image classifier. It is a Clinical Visual Intelligence Fabric (CVIF) capable of incorporating whichever model, algorithm or rule produces the most reliable evidence for a given intended use.
2. The Closed-Set Trap
Conventional supervised medical Vision AI generally follows a closed-set lifecycle: clinical images are annotated against a predefined taxonomy, a task-specific model is trained and validated, and production inference predicts among known classes.
The patient’s biology is not constrained by the model’s label schema.
A rare condition, a typical presentation, under-represented phenotype, novel artifact or clinically meaningful visual finding may sit outside the deployed taxonomy. A closed-set model is still useful, but it should not be mistaken for a complete representation of visual clinical reality.
AOV-CVI therefore separates two questions: (1) which validated known capability applies to this observation, and (2) what additional clinically meaningful evidence may be present beyond the known capability?
3. Traditional Computer Vision vs Open-Vocabulary Clinical Vision Intelligence
The distinction is not classifier versus Visual Language Models (VLM). It is Model-centric Vision AI versus evidence-centric Clinical Visual Intelligence.
4. Zero-Shot as Discovery, Not Diagnosis
Medical VLMs can align image and text representations so that an image can be evaluated against governed clinical concepts without training a new classification head for every exploratory concept. This changes the economics of discovery, semantic retrieval and rapid concept evaluation.
A simplified zero-shot pathway is: Zero-Shot Clinical Visual Evidence
MedSigLIP is an implementation example, not an architectural dependency. Its model documentation describes a medical image-text embedding space and intended application patterns including zero-shot classification and semantic image retrieval
However: Zero-shot capability ≠ Clinical accuracy ≠ Clinical validation ≠ Regulatory authorization.
Therefore the appropriate role of zero-shot intelligence is not:
Accordingly, AOV-CVI positions zero-shot inference as a discovery and evidence-expansion mechanism. Candidate concepts are subject to clinical review, semantic governance and validation before they become production knowledge.
5. From Classification to Evidence Generation
The primary architectural shift is from Image → Label toward Image/Video → Evidence Package. Evidence can be generated independently across global, localized, spatial, phenotype, normality, artifact, temporal and contextual channels.
5.1 Example: oral lesion screening
Instead of returning only a disease score, an evidence-first system may represent anatomy, lesion localization, morphology, color, texture, artifact status, deviation from normal references, semantic candidates, temporal change and reliability. Risk reasoning occurs after these evidence channels are assembled.
This separation improves traceability: a clinician can see what the system observed, what model or rule generated each observation, and which evidence drove the final disposition.
A classification-centric architecture might return: Leukoplakia: 0.78
An evidence-centric architecture might generate:
Only after evidence generation should risk reasoning occur.
6. The Governed Open Vocabulary
Open vocabulary must not mean uncontrolled vocabulary. AOV-CVI uses a governed clinical visual ontology to structure anatomy, visual findings, morphology, clinical conditions, differentials, normality, artifacts and temporal states.
Each governed concept should have a stable identifier, preferred term, synonyms, clinical definition, applicable anatomy and modality, relationships, prompt or concept templates where relevant, evidence requirements, provenance, version and validation status.
The vocabulary may be open for discovery; production clinical knowledge must be governed.
Where appropriate, semantic representations should align with established healthcare terminologies and interoperability structures rather than creating an isolated AI-only vocabulary.
7. Healthcare Vision AI Reference Architecture
Seven-Plane Clinical Visual Intelligence Fabric (CVIF)
Two vertical architectures operate across these planes: Agentic Clinical Vision Control Plane and MedOps & Governed Learning Plane.
7.1 Healthcare Visual World
X-ray, CT, MRI, ultrasound, pathology, ophthalmology, dermatology, endoscopy, surgery, dentistry, smartphone imaging, bedside video, robotics and wearable cameras.
7.2 Visual Acquisition & Edge Intelligence
Integrity, modality validation, anatomy relevance, acquisition quality, artifact detection, capture guidance, device context and privacy controls.
7.3 Visual Evidence Fabric
Images, video, frames, ROIs, masks, embeddings, phenotypes, spatial evidence, temporal evidence, coordinate mapping, provenance and lineage.
7.4 Memory, Semantics & Knowledge
Clinical Visual Memory, the Clinical Visual Semantic Fabric, knowledge graphs, Visual RAG, normal references and outcome-linked evidence.
7.5 Clinical Visual Intelligence Fabric (CVIF)
Specialist models, medical VLMs, zero/few-shot methods, detection, segmentation, phenotype extraction, normality models, video intelligence, temporal models and evidence fusion.
7.6 Safety, Reliability & Disposition
Calibration, OOD, uncertainty, conformal methods where appropriate, evidence agreement, abstention, human review, retake and escalation.
7.7 Clinical Experience & Action
Clinical review, decision support, search, alerts, workflow, referral, acquisition guidance and longitudinal visualization.
8. Clinical Visual Memory
Human clinicians rarely interpret a meaningful visual observation without history. AI systems often do. Clinical Visual Memory (CVM) addresses this mismatch by creating governed longitudinal memory across patients, lesions, anatomical regions, care episodes, procedures, populations and outcomes.
Visual memory turns AI from a stateless observer into a longitudinal evidence system.
The key architectural object becomes the Patient Visual Trajectory: a longitudinal representation of visual change rather than a collection of unrelated images.
9. Clinical Visual Semantic Fabric and Knowledge Graph
Pixels are not clinical knowledge. Visual evidence becomes more useful when it can be mapped to governed clinical concepts and connected to anatomy, diagnoses, pathology, procedures, treatment, outcomes and provenance.
Pixels to Computable Clinical Knowledge
An evidence graph could represent:
Lesion-123
├── located_in → Lateral Tongue
├── finding → White Patch
├── morphology → Heterogeneous
├── border → Irregular
├── temporal_change → Enlarging
├── resembles → Leukoplakia
├── differential → SCC
├── previous_state → Lesion-098
└── evidence_source → Image-567
A lesion node, for example, may be connected to anatomical location, visual finding, morphology, temporal state, candidate concepts, differentials, prior states and source images. This makes visual evidence queryable and reusable across clinical workflows.
FHIR ImagingStudy can expose significant DICOM study metadata while the actual DICOM instances remain in appropriate imaging storage; non-DICOM media require other resources or storage mechanisms. This reinforces an important architectural principle: interoperability metadata and image binaries need not be collapsed into the same store. So, interoperability requires it, these semantic representations should connect with established healthcare standards and terminology rather than creating an isolated AI ontology.
10. Clinical Visual RAG
Retrieval-Augmented Generation is commonly framed as document retrieval. Clinical Vision Intelligence extends retrieval to visual evidence. A current image or ROI can be encoded and used to retrieve prior patient images, similar validated cases, pathology-confirmed examples, outcome-linked cases or curated reference evidence.
Visual similarity does not establish clinical equivalence. Retrieval therefore provides evidence context, not diagnosis. Retrieved evidence should carry provenance, validation status, patient-context controls and intended-use constraints.
A Visual RAG system can retrieve clinically relevant visual evidence.
Similarity should never itself be interpreted as diagnosis.
Visual RAG provides evidence context, not automatic clinical equivalence.
11. Agentic Clinical Vision Control Plane (ACVCP)
As the number of evidence generators, memory systems, semantic services and safety checks grows, orchestration becomes a first-class concern. The Agentic Clinical Vision Control Plane coordinates these capabilities without granting agents unrestricted clinical authority.
Detailed Agent Catalogue
Every agent should operate through: Identity → Authorization → Permitted Tools → Policy → Evidence → Action → Audit
Agent autonomy should be proportional to clinical risk.
12. Video and Temporal Clinical Intelligence
Healthcare Vision AI cannot stop at images. Video introduces time: motion, events, state transitions, interactions and procedural phases. Longitudinal imaging introduces change across encounters.
Get Maruti Sivakumar V’s stories in your inbox
Join Medium for free to get updates from this writer.
Temporal intelligence changes the core question from ‘What is visible?’ to ‘What happened, what changed, and what is happening now?’ Applications include endoscopy, surgery, ultrasound, rehabilitation, gait, respiratory observation and procedural workflow.
The Patient Visual Trajectory unifies serial images and video-derived states into a governed longitudinal representation.
This creates opportunities across endoscopy, surgery, ultrasound, rehabilitation, gait, respiratory monitoring and other temporal visual domains.
The architectural object is no longer merely the image. It becomes the: Patient Visual Trajectory
13. Reliability, Uncertainty and Abstention
Open-vocabulary systems need a first-class reliability architecture because the set of candidate concepts can be broader than the validated closed-set taxonomy. Raw similarity must not be presented as clinical probability.
· Calibration — align model outputs with observed performance for the validated intended use.
· Acquisition OOD — identify images outside validated device, quality or capture distributions.
· Semantic OOD — identify observations that do not adequately correspond to the validated concept space.
· Evidence disagreement — preserve conflicts among global, localized, phenotype, temporal and multi-view evidence.
· Conformal methods — where validated appropriately, support prediction sets rather than forced single-class outputs.
· Selective prediction — permit automated output only when predefined reliability conditions are met.
· Abstention — explicitly return UNKNOWN, UNCERTAIN, RETAKE or HUMAN REVIEW when evidence is insufficient.
Current GMLP principles emphasize representative data, clinically relevant testing, human-AI team performance, deployed-model monitoring and management of retraining risks. Transparency principles likewise emphasize intended use, performance, limitations, logic where available, workflow context and lifecycle communication.
Raw image-text similarity is not inherently clinical confidence.
The architecture therefore supports:
This aligns conceptually with current regulatory emphasis on human-AI team performance, representative evaluation, monitoring and retraining-risk management in GMLP. (U.S. Food and Drug Administration)
Transparency guidance also emphasizes communicating intended use, performance, limitations and relevant logic to users — important considerations for any clinical evidence architecture. (U.S. Food and Drug Administration)
14. The Open-to-Known Clinical Learning (OKCL) Loop
AOV-CVI becomes adaptive through a governed learning loop. A previously unmodeled pattern may be surfaced by semantic OOD, repeated clinician escalation, deviation from normal references or open-vocabulary evidence. It is not automatically promoted to clinical truth.
This loop creates a new knowledge lifecycle: Unknown → Candidate → Qualified → Governed → Labeled → Validated → Known
The system can therefore maintain two complementary pathways: a known pathway using validated closed-set capabilities, and an open pathway that explores unfamiliar evidence under governance.
This is fundamentally different from uncontrolled continuous learning.
15. MedOps and the Human-Approval Architecture
MedOps is defined, specifically within this paradigm, as:
The governed operational discipline that transforms clinically qualified visual discoveries into validated, traceable, deployable and monitored healthcare AI capabilities.
Continuous learning must not be confused with uncontrolled self-modification. FDA’s January 2025 AI-enabled device lifecycle guidance remains draft and describes a total-product-lifecycle risk-management approach. FDA’s August 2025 final PCCP guidance addresses prospectively planned modifications to AI-enabled device software functions.
No zero-shot discovery should automatically trigger: Discovery → Training → Production
Instead, it should be: Discovery creates a candidate hypothesis, not a production truth.
This approach is consistent with lifecycle-oriented regulatory thinking. FDA’s January 2025 AI-enabled device guidance remains draft guidance and takes a total-product-lifecycle approach. FDA’s August 2025 final PCCP guidance addresses prospectively planned AI-enabled device modifications and continued assurance of safety and effectiveness. (U.S. Food and Drug Administration)
16. Clinical AI Configuration Bundle
Production behavior is determined by more than a model artifact. AOV-CVI therefore treats the Clinical AI Configuration Bundle as the unit of validation, deployment, versioning, monitoring, audit and rollback.
· Model(s) and model adapters
· Preprocessing profile
· Clinical concept ontology
· Prompt or concept library
· Concept embeddings
· Normal reference assets
· OOD configuration
· Calibration assets
· Conformal configuration where used
· Evidence-fusion logic
· Decision thresholds
· Safety and abstention policy
· Program or intended-use profile
An AI Dependency & Validation Graph should record dependencies among these assets. Replacing an encoder, for example, can require review or regeneration of concept embeddings, OOD references, calibration, conformal assets, fusion logic and threshold validation.
The bundle becomes the unit of: validation → deployment → versioning → monitoring → audit → rollback.
A corresponding AI Dependency & Validation Graph identifies what must be reconsidered when one component changes.
This provides a more realistic lifecycle model for modern Clinical AI than model-version tracking alone.
17. Specialty-Specific Scenarios
Scenario A — Oral Cancer Screening
A community health worker acquires several white-light smartphone images of the oral cavity. The system evaluates acquisition quality, anatomy and artifacts. A specialist detector identifies a suspicious lateral-tongue region. A medical VLM generates open-vocabulary semantic evidence. Phenotype analysis identifies an irregular, heterogeneous white lesion. Clinical Visual Memory reveals that the lesion has increased in area since a previous encounter. Rather than issuing a definitive diagnosis, CVIF produces a structured evidence package and risk/reliability assessment.
If evidence is sufficiently reliable, the workflow recommends the validated screening disposition. If semantic evidence is ambiguous, the Safety Agent abstains and escalates the case. A recurring visual phenotype not represented by the governed ontology can enter the Open-to-Known Learning Loop.
Scenario B — Dermatology
A patient presents serial smartphone images of a skin lesion. CVM associates the images with the same lesion across encounters. Temporal analysis identifies changes in border morphology and color distribution. Open-vocabulary semantic evaluation identifies candidate morphological concepts, while Visual RAG retrieves validated comparable cases.
The clinician sees not merely a classifier score but: current phenotype + change trajectory + semantic evidence + comparable evidence + uncertainty.
Scenario C — Radiology
A chest image produces a visual pattern not strongly represented by the validated closed-set classifier. Rather than forcing it into the nearest known category, semantic OOD is raised. The medical VLM evaluates governed radiological concepts while Visual RAG retrieves similar validated images. The radiologist receives an evidence package identifying the unusual visual characteristics and retrieved comparators. If repeated clinically confirmed cases reveal a meaningful underrepresented pattern, MedOps can initiate concept qualification.
Scenario D — Pathology
A whole-slide workflow identifies regions of interest using specialist pathology algorithms. Localized embeddings enable semantic retrieval against pathology-confirmed reference cases. Morphological concepts are connected through the Clinical Visual Semantic Fabric (CVSF) to relevant pathology terminology. Outcome-linked Visual Memory enables retrieval not merely by appearance but eventually by appearance + confirmed outcome.
Scenario E — Endoscopy
A video pipeline continuously evaluates frames and temporal segments. Specialist detectors identify known polyps. The temporal system tracks lesions across frames and avoids treating repeated views as independent discoveries. Open-vocabulary intelligence can characterize unusual visual patterns that do not map strongly to known categories.
The system can preserve the relevant video segment as evidence and escalate it for expert review.
Scenario F — Surgical Video
The system identifies anatomical structures, instruments and procedural phases. Temporal intelligence recognizes state transitions. The agentic control plane coordinates specialist anatomy models, instrument detectors and procedure-state models. Open-vocabulary reasoning is used for evidence enrichment — not unrestricted surgical decision-making. Unexpected events are preserved with temporal context for review and potentially later MedOps analysis.
Scenario G — Wound Care
Serial wound photographs become a Patient Visual Trajectory. The system measures:
- wound area
- tissue composition
- border changes
- color
- exudate-related visual characteristics where appropriately validated
- healing trajectory
CVM provides longitudinal comparison while the ontology represents wound-state concepts.
The important output becomes: how the wound is changing rather than simply: what class the current photograph belongs to.
Scenario H — Ophthalmology
Retinal images are compared against previous studies and validated normal reference distributions. Specialist retinal models provide known-condition evidence. A medical VLM provides complementary semantic evidence. Normality modelling can identify unusual patterns even where no governed disease concept provides a strong semantic match.
The system can therefore distinguish: “No known condition detected” from: “Evidence appears normal.”
18. Model-Agnostic Implementation
AOV-CVI deliberately avoids committing the architecture to one model family. Specialist supervised models, medical VLMs, normality models, temporal models and deterministic Computer Vision can coexist behind a governed model-adapter layer.
BiomedCLIP and MedSigLIP are useful implementation examples of biomedical/medical image-text representation models, but they are not the thesis of the architecture. The correct engineering choice depends on modality, intended use, compute envelope, validation evidence and failure modes.
The architecture deliberately avoids committing to one model family.
Choose models for evidence quality. Design the architecture for model change.
For a bounded screening task, a specialist classifier may be the strongest production capability. For open-vocabulary discovery, a medical VLM may be useful. For longitudinal video, a temporal model may be necessary. For some workflows, the strongest architecture may combine all of them.
19. Design Principles
The architecture can be summarized through ten principles:
- Evidence before prediction.
- Open for discovery; governed for production.
- Unknown is a valid answer.
- Risk and uncertainty are different concepts.
- Preserve source evidence and provenance.
- Memory should be longitudinal.
- Semantics turn pixels into knowledge.
- Agents orchestrate; policies authorize.
- Continuous learning does not mean uncontrolled self-learning.
- Govern the deployed capability, not merely the model.
20. Roadmap for Adoption
Organizations do not need to implement the complete paradigm at once.
A pragmatic maturity path is:
The destination should not dictate an unsafe leap from conventional CV directly to autonomous foundation-model reasoning.
The transformation should be incremental, validated and evidence-driven.
21. Conclusion
The future of Healthcare Vision AI is not a larger collection of image classifiers. Nor is it a single multimodal foundation model capable of answering every clinical question. It is an Adaptive Open-Vocabulary Clinical Vision Intelligence ecosystem.
Clinical Visual Intelligence Fabric is at its center and around it, Clinical Visual Memory, Clinical Visual Semantic Fabric, Knowledge Graphs, Visual RAG, specialist medical models, medical VLMs and temporal intelligence. Across it operates an Agentic Clinical Vision Control Plane. Around its decisions sits a Safety and Reliability architecture supporting calibration, OOD detection, uncertainty, evidence disagreement, selective prediction, abstention and human review. And around its lifecycle sits MedOps, implementing the Open-to-Known Clinical Learning Loop.
This changes the purpose of zero-shot Vision AI. Zero-shot learning is not the destination. It becomes a mechanism allowing healthcare AI to look beyond the boundaries of what it has already been explicitly taught to recognize.
The larger paradigm is: See → Understand → Remember → Relate → Reason → Verify → Act → Learn → Govern
And its operating contract can be stated even more simply:
The known is operationalized.
The unknown is surfaced.
The uncertain is escalated.
The new is clinically validated.
The validated becomes known.
Every transition remains governed.
That is the shift from Healthcare Computer Vision to Adaptive Clinical Visual Intelligence.
Key Propositions
Proposition 1 — Healthcare should move from classification-first to evidence-first Vision AI. The primary output of Vision AI should increasingly be structured clinical evidence rather than a single model label.
Proposition 2 — Open vocabulary should complement, not replace, validated closed-set AI. Known, validated clinical problems can continue to use specialist models. Open-vocabulary intelligence provides an additional mechanism for recognizing and investigating evidence outside those boundaries.
Proposition 3 — Zero-shot should be treated as discovery, not automatically as diagnosis. Zero-shot capability does not imply clinical validity, safety or regulatory authorization.
Proposition 4 — UNKNOWN is a valid clinical AI output. A clinically responsible architecture should support abstention, human review, retake and escalation rather than forcing every observation into a known category.
Proposition 5 — Images should become computable clinical knowledge. Visual findings should connect to anatomy, morphology, terminology, longitudinal history, clinical outcomes and provenance through ontologies and knowledge graphs.
Proposition 6 — Healthcare AI needs visual memory. Clinical interpretation should increasingly reason over patient visual trajectories, not isolated images.
Proposition 7 — Video requires temporal intelligence. Healthcare Vision AI must evolve from “What is visible?” toward “What happened, what changed and what is happening now?”
Proposition 8 — Agentic AI should orchestrate evidence, not possess unrestricted clinical authority. Agents should operate through identity, authorization, approved tools, policies and auditable evidence.
Proposition 9 — Open discovery requires a governed path to becoming known. A newly surfaced pattern should move through clinical qualification, ontology governance, annotation, validation and human approval before becoming a production capability.
Proposition 10 — The deployable unit is larger than the model. Models, ontologies, prompts, embeddings, calibration, OOD assets, thresholds and safety policies collectively determine behavior and should therefore be governed as a Clinical AI Configuration Bundle.
References
- International Medical Device Regulators Forum / U.S. FDA. Good Machine Learning Practice for Medical Device Development: Guiding Principles. The current FDA page notes the January 2025 IMDRF final document containing ten GMLP principles. (U.S. Food and Drug Administration)
FDA — Good Machine Learning Practice - U.S. FDA, Health Canada & MHRA. Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles. Addresses intended use, performance, limitations, human-AI interaction and lifecycle transparency. (U.S. Food and Drug Administration)
FDA — ML Medical Device Transparency Principles - U.S. FDA. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations. Draft guidance, January 2025; emphasizes total-product-lifecycle risk management for AI-enabled device software functions. (U.S. Food and Drug Administration)
FDA — AI-Enabled Device Lifecycle Draft Guidance - U.S. FDA. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions. Final guidance, August 2025; addresses controlled iterative modification of AI-enabled devices. (U.S. Food and Drug Administration)
FDA — Final PCCP Guidance - U.S. FDA, Health Canada & MHRA. Predetermined Change Control Plans for Machine Learning-Enabled Medical Devices: Guiding Principles. Describes planned modifications, modification protocols, impact assessment and TPLC considerations. (U.S. Food and Drug Administration)
FDA — PCCP Guiding Principles - Google Health AI Developer Foundations. MedSigLIP Model Card. Documents a medical image/text embedding model and intended applications including zero-shot classification and semantic image retrieval. (Google Developers)
Google — MedSigLIP Model Card - U.S. FDA. Artificial Intelligence-Enabled Medical Devices. Central FDA resource covering GMLP, PCCPs, transparency principles and AI-enabled medical-device resources. (U.S. Food and Drug Administration)
FDA — Artificial Intelligence-Enabled Medical Devices
