AI Stroke Detection on Brain Scans: Faster Treatment and New Recovery Tech

When a stroke strikes, every minute costs brain cells. AI software now reads emergency brain scans in minutes and alerts specialists, and new devices are helping some survivors regain arm movement.

The problem today: why every minute of a stroke matters

A stroke happens when blood stops reaching part of the brain. Most strokes are ischaemic, meaning a blood clot blocks an artery (a blood vessel carrying oxygen-rich blood). The rest are haemorrhagic, meaning a blood vessel bursts and bleeds into or around the brain. According to the World Stroke Organization's 2025 fact sheet, there are about 11.9 million new strokes in the world each year and about 7.25 million deaths from stroke. Almost 94 million people are living after a stroke, and about one in four adults over 25 will have a stroke at some point in their life. Around 65% of new strokes are caused by a clot, about 29% by bleeding inside the brain and about 6% by bleeding in the space around it.

Once a stroke has started, speed is everything. A widely quoted 2006 study in the journal Stroke estimated that, in a typical large clot-type stroke, the brain loses about 1.9 million nerve cells for every minute the blood supply is not restored. Doctors summarise this as "time is brain". The two main emergency treatments both aim to reopen the blocked artery:

• Thrombolysis: a clot-dissolving medicine given through a drip into a vein. UK guidance from NICE (the National Institute for Health and Care Excellence) says it should be given within 4.5 hours of the start of symptoms in suitable patients.
• Thrombectomy: a procedure in which a specialist doctor passes a thin tube (a catheter) through an artery, usually from the groin or wrist, up into the brain and pulls the clot out. It is used for clots in large brain arteries. NICE guidance allows it up to 24 hours after onset in selected people whose scans show that brain tissue can still be saved.

Thrombectomy is only done in specialist centres, and it depends on a chain of steps that can each add delay. Someone must read the scan, spot the blocked artery and call the stroke doctor, who then calls the thrombectomy team. Patients at a local hospital must be transferred by ambulance to a bigger centre. At night, a specialist scan reader may not be on site. This is where artificial intelligence (AI) software has found its first big role in stroke care. It does not treat the stroke. It helps the right doctors see the right scan sooner.

How AI stroke detection works

Stroke AI tools are computer programs trained on many thousands of brain scans that experts have already labelled. The programs learn patterns that point to a blocked artery, bleeding or damaged tissue. Once installed in a hospital, the software receives a copy of every relevant scan automatically, analyses it within minutes and sends results and alerts to the stroke team's phones and computers. A human doctor still looks at every scan and makes every decision.

Reading the plain CT scan

The first scan is usually a non-contrast CT (a quick X-ray based scan of the head, without dye). Its main job is to check for bleeding, because clot-dissolving medicine must never be given to someone bleeding in the brain. AI tools can flag possible bleeding. Some also produce an ASPECTS score, a 10-point score that estimates how much of a key brain area already shows early damage (a higher score means less damage). Different doctors can disagree when scoring by eye, so a consistent automatic number can help.

Finding the blocked artery

The second scan is often a CT angiogram (CTA), in which injected dye makes the blood vessels show up. AI software traces the vessels and looks for a sudden cut-off, which suggests a large vessel occlusion (a clot blocking a big artery), the kind thrombectomy can remove. It can alert the stroke team before a radiologist (a doctor who specialises in reading scans) has opened the images.

Measuring brain tissue that can still be saved

Some hospitals also do a CT perfusion scan, which follows the dye as it flows through the brain over time. From this, software creates coloured maps showing two areas. The core is tissue that has very likely already died. The penumbra is tissue that is short of blood but still alive and could recover if the artery is reopened. A large penumbra with a small core suggests treatment may still help, even many hours after the stroke began. Automated perfusion software of this kind was used to select patients in the DAWN trial (published in the New England Journal of Medicine in 2018), one of the studies that showed thrombectomy can help selected people up to 24 hours after symptoms start.

Sharing images across hospitals

Many stroke AI products include a secure phone app. A doctor at a local hospital can share the scan and AI result with the specialist at a thrombectomy centre in seconds, and the specialist can agree a transfer straight away. Faster communication may matter as much as the image analysis itself.

Technology for recovery: paired vagus nerve stimulation

Many survivors are left with weakness, often in one arm and hand, and progress in rehabilitation often slows after the first months. In paired vagus nerve stimulation (VNS), a small pacemaker-like device is placed under the skin of the chest, with a thin wire on the vagus nerve in the neck. Each time the person does a movement exercise in therapy, the device sends a brief, mild pulse to the nerve. This is thought to release brain chemicals that help strengthen the connections being used, supporting neuroplasticity (the brain's ability to rewire itself).

AI-assisted rehabilitation and other research tools

Other tools are being studied: robots that help move a weak limb and adjust their help to the patient's own effort, and brain-computer interfaces, which read brain signals through a cap of sensors and use them to move a hand device. A non-surgical form of vagus nerve stimulation through the ear is also being tested; a small 2021 pilot in Frontiers in Neuroscience combined it with robotic arm training and reported less muscle stiffness. These remain mostly at the research stage.

The main tools and where they stand

Several companies make stroke imaging AI. The products below are the ones with the clearest public regulatory record. Approval in one country does not mean approval everywhere, and products are updated often, so this is a snapshot.

Viz.ai (Viz LVO, earlier called ContaCT)

Made by Viz.ai (United States). It detects large vessel blockages on CT angiograms and alerts the team by phone. In February 2018 the US Food and Drug Administration (FDA) authorised it through its De Novo route for new types of device, making it one of the first AI triage tools in medicine. NICE added Viz to its NHS list in May 2024.

Brainomix 360 Stroke (e-Stroke)

Made by Brainomix (Oxford, UK). It offers automatic ASPECTS scoring (e-ASPECTS), blocked vessel detection, perfusion analysis and image sharing. e-ASPECTS received FDA clearance in March 2023. It was the system studied in the large English evaluation described below.

RapidAI

Made by RapidAI, formerly iSchemaView (United States). Its modules cover perfusion maps, ASPECTS, bleeding and blockage detection. Rapid LVO received FDA clearance in August 2020.

Aidoc and Avicenna.AI

Aidoc (Israel) makes BriefCase tools that flag suspected bleeding (FDA clearance reported in August 2018) and large vessel blockage (cleared December 2019, FDA file K192383). Avicenna.AI (France) makes CINA tools, and its CINA-ASPECTS received FDA clearance in April 2024. NICE placed both among tools needing more research before routine NHS use.

Vivistim Paired VNS System

Made by MicroTransponder (United States). The FDA approved it on 27 August 2021, with Breakthrough Device designation, for moderate to severe arm and hand weakness after a past ischaemic stroke (a stroke caused by a clot).

NICE notes that the tools recommended for NHS use should only be used once they also have the NHS's digital technology safety approval (called DTAC). The other tools NICE reviewed, including Aidoc and CINA Head, should for now only be used in research.

What the evidence shows for AI stroke software

A randomised trial in Houston, United States

The strongest test so far is a UTHealth Houston trial published in JAMA Neurology in 2023. It included 243 thrombectomy patients at four stroke centres between 2021 and 2022. Each hospital switched to AI detection with phone alerts at a randomly chosen time (a stepped-wedge cluster trial). With AI, the time from arrival to the start of thrombectomy fell by about 11.2 minutes. After adjusting for age and other factors, there was no difference in how many patients could live independently afterwards. The software made care faster, but the trial did not show better recovery.

The English national rollout

England has used stroke AI more widely than almost any other country. NHS England said in December 2024 that all 107 stroke centres in England were using AI to help assess brain scans for around 80,000 people with stroke each year. A large study of this rollout, led by researchers in Reading and Oxford and published in The Lancet Digital Health in late 2025, used national stroke audit data covering 452,952 stroke admissions between 2019 and 2023. It focused on 26 hospitals in four regional networks that introduced the Brainomix system. The study found:

• At these hospitals, the share of stroke patients who had thrombectomy rose from 2.3% before the software to 4.6% after it. At comparison hospitals the share also rose, from 1.6% to 2.6%, which shows that thrombectomy services were growing everywhere during this period.
• Among 747 patients transferred from a local hospital to a thrombectomy centre, the median time from arriving at the first hospital to leaving it (called "door-in-door-out" time) was 128 minutes when AI was used and 192 minutes when it was not.
• Patients whose scans were reviewed with AI had higher odds of receiving thrombectomy than those whose scans were not.

This was an observational study, so other changes at the same time, such as extra staff or thrombectomy capacity, could explain part of the improvement. NHS England also reported early figures suggesting the share of patients left with no or only slight disability rose from 16% to 48% after AI was introduced. These early observational numbers need caution, especially as the Houston randomised trial showed no difference in recovery.

Independent reviews

A 2024 systematic review for the UK's National Institute for Health and Care Research (NIHR) found that most studies measured how accurately the software reads scans, with much less evidence about what happens to patients. NICE reached a similar view, which is why it recommended the main tools only alongside further evidence collection.

Evidence for vagus nerve stimulation

The main trial behind Vivistim is VNS-REHAB (ClinicalTrials.gov NCT03131960), published in The Lancet in April 2021. It included 108 people with moderate to severe arm weakness after an ischaemic stroke, at 19 sites in the US and UK. All had six weeks of intensive arm rehabilitation in clinic, followed by home exercises. About half received real stimulation during exercises and the rest received sham (pretend) stimulation, and neither patients nor assessors knew which. Arm movement was measured with the Fugl-Meyer upper limb score, a standard test of arm function.

• At 90 days, the stimulation group improved by an average of 5.0 points, compared with 2.4 points in the sham group.
• 47% of the stimulation group improved by 6 points or more, a change most patients would notice, compared with 24% of the sham group.

A one-year follow-up of 74 participants, published in the journal Stroke, reported an average gain of about 5.2 points from the start of the trial. These are meaningful but modest changes. Stimulation helped some people noticeably, while others gained little.

Who can benefit from AI stroke software

Stroke imaging AI

There is no eligibility test. Where a hospital has the software, it is usually applied automatically to everyone scanned for a suspected stroke. It is most useful for:

• People with a suspected clot in a large brain artery, who may need thrombectomy.
• People who arrive at a local hospital that would need to transfer them to a specialist centre.
• Hospitals where a specialist scan reader is not always on site, such as at night.

It helps less for small strokes deep in the brain, strokes in the back of the brain, and mini-strokes (transient ischaemic attacks or TIAs) where symptoms go away and scans may look normal. In these cases the doctor's examination and specialist reading matter most.

Vagus nerve stimulation

Based on the FDA approval and the main trial, paired VNS is aimed at adults who:

• Had an ischaemic (clot-type) stroke some time ago. In the trial, participants were between nine months and ten years after their stroke.
• Still have moderate to severe weakness in one arm and hand.
• Have some movement left to build on, and can commit to an intensive exercise programme and home practice.

It is generally not for people who had a bleeding stroke (they were not studied), people in the first months after a stroke, or people not fit for an operation.

What to expect as a patient

In the emergency department

1. Call and ambulance. Paramedics often alert the hospital that a possible stroke patient is on the way.
2. Quick assessment. A doctor checks speech, face and limb strength. The time the person was last seen well is vital, so family members should share it.
3. Scans. A plain CT, usually followed by a CT angiogram and sometimes a perfusion scan. These take only minutes.
4. AI analysis in the background. The software sends results to the team. Patients notice nothing different.
5. Decision. Doctors combine scans, their own reading and the examination. Clot-dissolving medicine may be started, and the thrombectomy team called or a transfer arranged.
6. Thrombectomy, if needed. Done in a special X-ray room under sedation or general anaesthetic, followed by monitoring on a stroke unit.
7. Recovery. A hospital stay of days or longer, then therapy at home or in a rehabilitation unit.

If you are considering vagus nerve stimulation

1. Referral and tests. A neurologist or rehabilitation doctor checks whether you meet the criteria, assesses your arm and checks your general health.
2. The operation. A surgeon places the device under the skin of the upper chest and attaches the lead to the vagus nerve through a small cut in the neck.
3. Clinic therapy. After healing, intensive arm therapy with stimulation paired to each movement. In the trial this lasted six weeks.
4. Home therapy and follow-up. Daily exercises at home, with regular checks of the device and your progress.

AI stroke software: access and cost in the US, UK and EU

United States

The imaging AI tools above are FDA authorised and used in many US hospitals. Patients do not buy the AI; it is part of the hospital's stroke service, and there is no official data showing a separate patient charge. Vivistim is offered at a number of US centres. Insurance coverage for the device and surgery varies, so check your plan before treatment.

United Kingdom

Stroke imaging AI is used in all 107 stroke centres in England, according to NHS England, and is part of NHS care at no extra cost to patients. NICE recommends three tools (e-Stroke, RapidAI and Viz) for NHS use while more evidence is collected. Part of the main paired vagus nerve stimulation trial took place in the UK, but there is no official data on routine NHS availability. Ask your stroke or neurology team whether any research studies are open.

European Union

Several of these tools carry the CE mark needed for sale in Europe. Use varies between countries and hospitals, and there is no official EU-wide data on coverage. Patients in public systems do not usually pay for the AI directly. There is no official data on the EU status of Vivistim.

Elsewhere and for international patients

For other countries there is no official data in the sources reviewed here. Emergency stroke care must happen where the stroke happens. For planned treatments such as vagus nerve stimulation, if you consider travelling, check:

• Approval: is the device approved by the regulator in that country, and for your type of stroke?
• Experience: how many patients has the centre treated, and does it have a dedicated rehabilitation team trained in the method?
• Records: bring your scans on disc, discharge letters, a list of medicines and a recent therapy assessment.
• Aftercare: who will check and adjust the device when you return home, and can a team in your home country support it?
• Insurance and total cost: get a written quote that includes surgery, therapy, travel and follow-up, and check whether your insurer will cover any of it.

Risks, limits and realistic expectations

• AI can miss strokes. It is best at clots in the largest arteries at the front of the brain, and less reliable for smaller branches, strokes at the back of the brain and very small strokes. "No blockage detected" does not mean there is no stroke, which is why a specialist must still read every scan.
• False alarms. Narrowed arteries from older disease, unusual anatomy, dental metal or poorly timed dye can trigger alerts. A doctor checks every alert before acting.
• Over-reliance. Busy doctors could trust the software too much or lose scan-reading skills over time.
• Technical failures. Patient movement or unusual scan settings can stop the software working. NICE lists real-world failure rates as an evidence gap.
• Faster is not proven better. Speed and organisation improve, but better recovery has not been proven in a randomised trial.
• Vagus nerve stimulation needs an operation, with risks such as infection, and hoarseness can occur because the vagus nerve lies near the nerves of the voice box. Gains were noticeable for about half of trial patients, need hard work in therapy, and do not reverse the stroke.

Clinical trials and what is next for stroke

NICE's evidence generation plan asks for real-world data on treatment times, thrombectomy rates, software failures and everyday accuracy for the three tools it recommends. Researchers are also moving towards larger studies of what happens to patients. Areas being explored include:

• Better detection of smaller and harder-to-see clots, including those at the back of the brain.
• Software that predicts a patient's likely recovery to support treatment decisions. These tools are still being tested and are not approved for this purpose.
• Non-surgical ear-based vagus nerve stimulation combined with therapy or robotics.
• AI-driven rehabilitation robots and brain-computer interfaces that respond to the patient's own effort.

To look for trials you may be able to join, search ClinicalTrials.gov (worldwide), the EU Clinical Trials Information System, or the UK's Be Part of Research service, and ask your stroke team. Joining a trial is voluntary, and you can leave at any time.

Questions to ask your doctor about AI stroke software

These questions may help you understand the care you or your family member received, or plan the next step.

• Does this hospital use AI software to help read stroke scans? Which one?
• What did the scans show: a clot, bleeding, or neither? Was a large artery blocked?
• Was clot-dissolving medicine or thrombectomy considered, and why was it given or not given?
• If transfer to another hospital was needed, how long did it take?
• What is causing my remaining symptoms, and what is the rehabilitation plan?
• Am I likely to keep improving, and over what time frame?
• Am I a candidate for paired vagus nerve stimulation or any other approved recovery device?
• Are there clinical trials of new rehabilitation methods that I could join?
• What can I do to lower my risk of another stroke (blood pressure, medicines, smoking, activity)?
• Who should I contact if new symptoms appear after I leave hospital?

If new stroke symptoms appear at any time, do not wait for an appointment. Call emergency services immediately.

Questions patients ask