Executive summary
In 2025, the use of AI in medical diagnostics within the NHS has moved decisively from research concepts to regulated, real-world clinical practice. A growing number of proven tools are already being deployed across key pathways, including radiology, dermatology, ophthalmology, endoscopy, cardiology, and pathology. Early results from these deployments show the tangible benefits of AI: faster and more accurate pathways, higher detection rates for critical diseases, and significant efficiency gains (PubMed, GOV.UK, NICE).
However, this adoption is not a free-for-all. It is governed by a robust and maturing UK assurance framework. Any AI for medical diagnosis in the NHS must navigate the rules for medical devices set by the MHRA, meet the evidence standards of NICE (often via an Early Value Assessment), and pass the baseline procurement requirements of the DTAC. This guide provides a clinician-focused map of the diagnostic AI that is deployable in the NHS today, what’s in the pipeline, and the essential governance context for safe adoption.
The landscape: why “diagnostic AI” is not one thing
The term "diagnostic AI" covers a broad spectrum of technologies, which can be grouped into three main categories:
- Perception AI: This is the most mature category and involves the interpretation of images or physiological signals. It is dominant in radiology, endoscopy, dermatology, ophthalmology, and ECG analysis.
- Reasoning AI: These are tools designed to assist with differential diagnosis and clinical triage by synthesising complex patient data.
- Predictive risk AI: These models analyse datasets to predict patient deterioration (e.g., sepsis risk) or disease prognosis.
What clinicians can use now (NHS/UK focus), by specialty
Radiology & imaging
- Hyperacute stroke: Brainomix e-Stroke is widely deployed across NHS networks, with reports showing significant improvements in process times for thrombectomy decisions.
- Fractures on X-ray (urgent care): Under NICE's EVA pathway (HTE20), AI tools like BoneView, Rayvolve, RBfracture, and TechCare Alert can be used to support the detection of fractures in urgent care settings while further evidence is generated.
- Chest X-ray (CXR) assistance: Annalise Enterprise CXR has been rolled out across multiple English imaging networks, funded via the AI Diagnostic Fund, with real-world evaluations describing its impact.
- Breast screening: A large-scale UK trial is underway. In parallel, UK sites using tools like Kheiron’s Mia® in prospective evaluations have reported both workload reduction and an uplift in cancer detection.
Gastroenterology (endoscopy)
- Polyp/adenoma detection: The GI Genius module from Medtronic, which uses AI to highlight suspicious polyps during a colonoscopy, is supported by a wealth of evidence showing it increases the adenoma detection rate (ADR) and reduces miss rates.
Dermatology
- Skin cancer pathways: Skin Analytics’ DERM is a CE-marked (Class IIa) device that is being deployed in NHS trusts to help triage urgent skin cancer referrals. It has a conditional recommendation from NICE, enabling its use while further real-world evidence is gathered.
Ophthalmology
- Autonomous diabetic retinopathy screening: Tools like IDx-DR, EyeArt, and AEYE are FDA-cleared in the US, and their mature international evidence base is informing their adoption path in the UK.
Cardiology
- Non-invasive coronary physiology: HeartFlow FFRct has a long-standing positive recommendation from NICE (MTG32). It uses AI to analyse CT coronary angiograms, providing a cost-saving, non-invasive alternative to traditional functional testing.
Pathology
- Prostate & breast cancer support: Ibex Medical Analytics has announced expanding deployments across UK trusts. Paige, another leader in the field, has a suite of CE-marked and FDA-cleared products, including a recent FDA Breakthrough designation for its Pan-Cancer Detect tool, with evaluations running in major NHS labs.
What’s emerging (pilots & research you should watch)
- Diagnostic dialogue systems: Google’s AMIE has shown promise in clinician-level diagnostic dialogue on complex vignettes in research published in Nature. In a real-world setting, DxGPT has been piloted by Madrid’s public health system to support rare disease diagnosis.
- Foundation models for imaging: RETFound, a UK-led foundation model for ophthalmology from Moorfields and UCL, is demonstrating how large-scale models can be adapted efficiently to new diagnostic tasks.
- Predictive/prognostic AI: While the literature on sepsis and deterioration prediction is mixed (with some well-known models performing poorly on external validation), the field is advancing rapidly, with a focus on more robust, locally-tuned models.
Safety, evidence & governance (UK specifics)
- NICE Early Value Assessment (EVA): This is the key pathway for getting promising but early-stage AI into the NHS. It allows for conditional adoption with a clear requirement to generate real-world evidence.
- MHRA AI Airlock: This regulatory sandbox, now in its second phase, is accelerating learning for novel and adaptive AIaMD, helping to shape future regulations.
- Operational playbooks: NHS England's AI Knowledge Repository is becoming the national hub for sharing lessons and best practices from real-world evaluations, including those from the AI in Health and Care Award.
Practical buyer’s guide: selecting a diagnostic AI
- Clinical value: Is there peer-reviewed evidence (RCT or real-world) for the tool? Does it improve a clinically relevant endpoint (e.g., adenoma detection rate, time-to-thrombectomy)?
- Workflow fit: How does it integrate with your PACS or EHR? Is it designed for human-in-the-loop use or as a fully autonomous tool?
- Assurance pack: Can the vendor provide their DTAC and DCB0129/0160 safety cases? Have they aligned their evidence plan with the NICE EVA framework?
How reasoning-assist tools fit day-to-day
While the tools above focus on perception AI, a different class of reasoning-assist tools can help clinicians synthesise information.
- Cited clinical search & Q&A: Tools like OpenEvidence (in the US) and the UK-centric iatroX can help to close knowledge gaps by providing citation-first answers to clinical questions.
- Differential “brainstorming”: Tools like iatroX Brainstorm or DxGPT can be used as a "second pair of eyes" to broaden a differential diagnosis in a complex case, but they are not a substitute for a final arbiter.
Quick comparison table
| Use-case | NICE Status | Setting | Oversight Required |
|---|---|---|---|
| Fracture detection (X-ray) | EVA HTE20 | Urgent Care/ED | Clinician review mandatory |
| Skin-lesion triage (2WW) | EVA HTE24 | Dermatology Referral | Within EVA protocol |
| Coronary physiology (CCTA) | MTG32 | Cardiology/Imaging | MDT use |
| Stroke decision-support | DG57 (Evidence-generating) | Stroke Networks | Specialist oversight mandatory |
FAQs
- Are AI “diagnosis” tools allowed in the NHS?
- Yes, but only when they are properly regulated as a medical device by the MHRA, have a sufficient evidence base as assessed by NICE (often via an EVA), and are deployed with robust local governance (DCB/DTAC).
- What’s the fastest route for a new AI diagnostic tool into practice?
- The NICE EVA pathway is specifically designed for this, enabling time-limited, conditional use while the required real-world evidence is generated.
- Where can NHS teams get end-to-end guidance on adopting AI?
- The AI & Digital Regulations Service (AIDRS) portal is the official "one-stop shop" for both adopters and developers.