How accurate is AI at detecting intracranial haemorrhage on CT scans?

Studies reviewed report sensitivity and specificity often above 90%, but most figures come from retrospective curated datasets, which tend to overestimate real-world performance.

What are the main limitations of AI-based brain bleed detection?

Key limitations include distribution shift between training data and real clinical environments, variability in haemorrhage subtypes, small study sample sizes, and lack of prospective multi-site validation.

Is AI ready to replace radiologists in detecting brain bleeds?

No. Current AI tools are best positioned as triage flags or decision-support systems to prioritize urgent scans, not as autonomous diagnostic replacements for trained radiologists.

What should hospitals ask before buying an AI radiology triage tool?

Buyers should ask about training data distribution, performance by haemorrhage subtype, prospective validation results, false negative rates at operational thresholds, and regulatory classification under frameworks like FDA SaMD.

Home›Computer Vision›Article

Computer Vision

Can AI Actually Read CT Scans for Brain Bleeds? A Systematic Review Breaks It Down

A systematic review in Cureus examines how accurately AI detects intracranial haemorrhage on non-contrast CT scans — and finds impressive headline numbers that come with important real-world caveats about dataset quality and deployment gaps.

Dispatch desk·June 28, 2026·4 min read

TL;DR

AI models show high sensitivity and specificity for brain bleed detection on CT scans, but most evidence comes from curated retrospective datasets that don't reflect messy real-world clinical environments.

What changed

A systematic review published in Cureus aggregated and critically assessed the diagnostic accuracy of AI systems for detecting intracranial haemorrhage on non-contrast CT head scans across multiple studies.

Why it matters

Clinicians and health tech buyers need to understand the gap between benchmark accuracy and real-world performance before deploying AI triage tools in high-stakes emergency radiology workflows.

Editorial record

Drafted by autopilot. Reviewed when marked.

Source

Intracranial haemorrhage — bleeding inside the skull — is the kind of diagnosis where minutes genuinely matter. Miss it on a CT scan, and a patient's outcome can deteriorate catastrophically. So when AI vendors pitch their computer vision tools as reliable detectors of brain bleeds on non-contrast CT heads, the stakes aren't abstract. They're life-and-death concrete. That's exactly why a systematic review published in Cureus set out to rigorously examine just how accurate these AI systems actually are — and the findings are worth unpacking carefully.

The Setup: What Are We Actually Measuring?

Non-contrast CT of the head (NCCT) is the standard first-line imaging tool in emergency settings when intracranial haemorrhage (ICH) is suspected. It's fast, widely available, and doesn't require contrast injection. The catch? Reading it accurately under pressure, at 3 a.m., in a busy emergency department, is genuinely hard. Subtle bleeds can be missed. Radiologist fatigue is real. Enter AI-based triage and detection tools, promising to flag critical findings before a human even touches the scan.

The systematic review focused on the diagnostic accuracy of AI algorithms — primarily deep learning-based computer vision models — in detecting ICH on NCCT. The key metrics evaluated were sensitivity (how often the AI correctly identifies a bleed when one exists) and specificity (how often it correctly clears a scan when no bleed is present). Get the sensitivity wrong and you're missing bleeds. Get the specificity wrong and you're flooding radiologists with false alarms — which, at scale, creates its own crisis of alert fatigue.

What the Evidence Actually Shows

Across the studies included in the review, AI models demonstrated generally high sensitivity and specificity for ICH detection — numbers that, on the surface, look impressive. Several systems reported sensitivity values north of 90%, with some approaching 95% or higher under controlled conditions. Specificity figures followed a similar pattern.

Before you get too excited, here's the important asterisk: a lot of this performance data comes from retrospective studies run on curated datasets. That's benchmark theater territory. Curated datasets tend to be cleaner, better-labelled, and more balanced than the chaotic mix of scan quality, patient demographics, and edge-case pathologies that show up in real emergency departments. Models trained and tested on pristine academic datasets frequently stumble when deployed in the wild — a phenomenon well-known in machine learning circles as distribution shift.

The review also highlighted variability across studies in terms of AI architecture, training data sources, reference standards used for ground truth, and the specific subtypes of haemorrhage being detected. Subdural haematomas, epidural bleeds, subarachnoid haemorrhages, intraparenchymal bleeds — each has different imaging characteristics, and performance varied accordingly. An AI that's excellent at catching the dramatic hyperdensity of a large intraparenchymal bleed may struggle with a thin, subtle subdural collection.

The Methodological Minefield

Here's where any honest engineer has to slow down. Systematic reviews of diagnostic AI are tricky because the underlying studies themselves often have methodological inconsistencies. The Cureus review flagged issues that are frustratingly common in this space: heterogeneous study designs, lack of prospective validation, small sample sizes, and insufficient reporting of confidence intervals around accuracy metrics. A sensitivity of 93% sounds great until you realize the confidence interval spans from 87% to 97% — that's a clinically meaningful range of uncertainty.

There's also the question of what "detection" actually means in these studies. Some AI tools are designed as triage flags — they raise an alert to prioritize radiologist review, rather than making a standalone diagnosis. Others are positioned as decision-support tools. Conflating the two when measuring accuracy leads to apples-and-oranges comparisons that muddy the evidence base considerably.

Why This Still Matters (A Lot)

None of the above is a reason to dismiss AI-assisted ICH detection. Far from it. The clinical case for AI triage in radiology is legitimately compelling: radiologist shortages are real, scan volumes are rising, and the window for intervention in ICH is narrow. A tool that reliably surfaces high-priority scans for faster review — even if it's not perfect — has genuine value in a well-designed workflow.

The smarter question isn't "is AI accurate?" in the abstract. It's "accurate enough, for which specific task, in which deployment environment, with what safeguards?" A high-sensitivity triage flag that occasionally over-alerts is a very different product from an autonomous diagnostic system. Regulatory frameworks like the FDA's Software as a Medical Device (SaMD) pathway are slowly catching up to these distinctions, but the clinical literature hasn't always been rigorous about making them explicit.

What Builders and Buyers Should Take Away

If you're building AI diagnostics for radiology, the systematic review is a useful reality check. Retrospective accuracy on benchmark datasets is a necessary starting point — not a finishing line. Prospective, multi-site clinical validation with diverse patient populations and real-world scan variability is what separates a publishable proof-of-concept from something you can actually deploy responsibly in a hospital.

If you're evaluating AI tools for clinical procurement, don't let vendors lead with aggregate sensitivity/specificity figures without context. Ask about the training data distribution, the case mix in validation cohorts, performance stratified by haemorrhage subtype, and — critically — what the false negative rate looks like at your target operating threshold. That last number is the one that should keep you up at night.

The bottom line from this systematic review is cautiously optimistic: AI shows real promise for ICH detection on NCCT, but the evidence base needs more prospective, real-world validation before we can confidently map headline accuracy numbers onto clinical outcomes. The technology is capable. The question is whether the deployment science is keeping pace.