Most AI tools built to catch stroke in a hospital’s imaging queue are trained to read the scan that comes second: the CT angiogram, taken after a radiologist or ER doctor already suspects a large vessel occlusion. On March 5, 2026, Sydney- and Boston-based Harrison.ai announced FDA 510(k) clearance for a tool that instead reads the scan that comes first — the plain, non-contrast CT brain scan every suspected stroke patient gets within minutes of arriving at the ER, before anyone has decided they need the more specialized, contrast-dependent scan at all.
Why the First Scan Matters More Than It Sounds
Non-contrast CT is the default opening move in stroke workups because it is fast and doesn’t require injecting contrast dye, which takes extra time and carries its own risks. CT angiography (CTA), the scan most existing stroke-AI products are built around, typically comes afterward, once a clinician has already decided a patient looks like a stroke case. Harrison.ai’s pitch is that triaging on the very first image — rather than waiting for someone to order the second one — can shave meaningful time off a process where, as the company’s co-founder and CEO Dr. Aengus Tran put it, “time is brain.” Patients who aren’t immediately placed on a dedicated stroke protocol, he said, can face significant delays before an infarct is even identified.
What the Tool Actually Does
The newly cleared product flags suspected acute ischemic infarcts across six vascular territories and mechanisms: the anterior, middle, and posterior cerebral arteries, plus cerebellar, basilar, and watershed infarcts. That’s a broader net than most competing tools, which generally target large vessel occlusion (LVO) in one or two territories and only on CTA. Harrison.ai’s chief medical officer, Dr. Jarrel Seah, called acute infarct detection on non-contrast CT “among the hardest findings in radiology” and said he expects most radiologists to welcome a tool that helps move suspicious cases up the reading queue.
The Numbers Behind the Clearance
In its FDA submission, the tool achieved up to 89.2% sensitivity on thin-slice CT images and 85.7% on thick-slice images, with greater than 80% sensitivity and specificity at multiple operating points on both. The reference standard for confirming true infarcts was advanced imaging, and Harrison.ai noted the infarcts used for validation weren’t guaranteed to be visible on non-contrast CT in the first place — a built-in reminder that the comparison is somewhat generous to the AI. The company also benchmarked itself against an existing FDA-cleared non-contrast CT LVO-only triage device, which posted 63.5% sensitivity and 95.1% specificity for vessel occlusion alone, and against a CTA-based LVO reference device that hit 91.3% sensitivity and 85.6% specificity — but only on the later, contrast-dependent scan.
A Breakthrough Device, Repeatedly
This is Harrison.ai’s ninth FDA clearance overall, covering thirteen radiological findings across nine separate 510(k) submissions. Three of those indications carry the FDA’s Breakthrough Device Designation, a status meant to speed review for technologies addressing unmet medical needs; the company says its acute infarct and obstructive hydrocephalus triage tools are among only about 15% of Breakthrough-designated devices that have gone on to actually receive marketing authorization. Combined with seven previously cleared non-contrast CT brain findings, Harrison.ai is positioning itself as having the broadest single-platform coverage of critical head CT findings of any AI triage vendor on the market.
The Skeptical Read
Clearance is not the same as proof of better patient outcomes. The FDA’s 510(k) pathway establishes that a device is safe and performs comparably to existing tools, not that it reduces time-to-treatment or improves survival in real hospital workflows — those outcome studies, if they come, would arrive later. There’s also the honest caveat buried in Harrison.ai’s own data: because validation relied on infarcts confirmed by advanced imaging that isn’t always visible on the non-contrast scan the AI actually reads, real-world sensitivity in a messier, more varied patient population could land lower than the submission numbers suggest. And a tool that flags more potential infarcts across six territories, rather than one or two, inevitably raises questions about false-positive rates and how radiologists will be expected to triage AI alerts without simply adding noise to an already overloaded reading list.
What Happens Next
The practical test will be adoption: whether hospitals running non-contrast CT as their first stroke scan start routing it through Harrison.ai’s model before a clinician even orders the angiogram, and whether that earlier flag genuinely compresses door-to-treatment time for thrombectomy or clot-busting drugs, where minutes are tied directly to how much brain tissue survives. Harrison.ai still has a third Breakthrough-designated product awaiting its own clearance, and competitors building LVO-only tools are likely to respond by trying to extend their own coverage beyond vessel occlusion. For now, the clearance marks a shift in where AI is being pointed in the stroke pathway — earlier, onto the scan everyone already gets, rather than the one only some patients end up needing.
Photo: Engin_Akyurt / PIXABAY via Pixabay
