A brain-aneurysm detector that scores 95% sensitivity on the cohort it was trained on can still be clinically useless. The number that looks like success is measured on the wrong axis. Aneurysm detection is not really an object-detection problem you can close with one confident sensitivity figure — it is a 3D segmentation-plus-detection pipeline whose viability lives or dies on three things the headline number hides: how well it finds small lesions, how many false alarms it raises per scan, and whether any of it survives a different scanner in a different hospital.

That gap between what a detector reports and what a radiology department will actually tolerate is where most aneurysm-detection programmes stall. Understanding why requires separating the parts of the problem that look identical to a generic detector but behave nothing alike in a clinical reading room.

How does aneurysm detection work in practice?

An intracranial aneurysm is a localised bulge in a cerebral artery wall. On imaging it shows up on CT angiography (CTA) or MR angiography (MRA) as a small outpouching on a vessel — often just a few millimetres across, embedded in a dense tangle of overlapping vasculature. The clinical task is computer-aided detection (CADe): flag candidate aneurysms so a radiologist reviews them, without replacing the radiologist’s judgment.

In practice the pipeline is rarely a single detector. A common pattern is to first segment the vascular tree from the volumetric scan, then run detection on that vessel geometry rather than on raw voxels, so the model reasons about candidates in the context of the vessel they sit on. This matters because the naive framing — “train a 3D bounding-box detector, report sensitivity, ship it” — treats aneurysm detection as if it were the same task as finding cars in street photos. It is not. The object is tiny, the background is anatomically complex, and the cost of a false positive is a radiologist’s attention, which is the scarcest resource in the workflow.

The distinction between raw detection accuracy and clinically usable accuracy is the whole game here. It is the same theme we develop in aneurysm detection with computer vision and how it works in practice; this article stays on why the naive framing misleads rather than the step-by-step pipeline.

Which CV patterns power aneurysm detection — 3D segmentation, detection, or a combined CADe pipeline?

All three, and the interesting engineering decisions are about how they combine. A CTA or MRA study is a 3D volume, so the natural backbone is a 3D convolutional network (3D U-Net variants are a standard starting point for the segmentation stage) rather than the 2D CNNs that dominate photographic vision. Promptable segmentation models — the medical adaptations of the Segment Anything Model in medical imaging — have made the vessel-and-lesion segmentation stage more data-efficient, though clinical use still demands careful validation rather than zero-shot trust.

The layering typically looks like this:

• Vessel segmentation isolates the arterial tree, reducing the search space and suppressing anatomically implausible candidates.
• Candidate detection proposes locations that look like outpouchings, tuned for high recall — it is cheaper to over-propose here and filter later.
• False-positive reduction (often a separate classifier operating on each candidate) is where most of the clinical value is won or lost.

The reason to treat this as a pipeline rather than one end-to-end detector is diagnostic transparency. When a combined CADe system misses a lesion or floods a scan with false marks, you need to know which stage failed. A monolithic model gives you a single number and no way to reason about the failure mode — the exact opposite of what you want when the output is decision-grade evidence for a regulator.

Why is sub-5mm sensitivity the hardest part of aneurysm detection, and how is it measured?

Aneurysm risk does not scale neatly with size, but detectability does. Large aneurysms are relatively easy to find; the small ones — particularly under 5mm — are where detectors quietly fail, and where a single aggregate sensitivity number hides the failure.

The correct measurement is per-lesion sensitivity stratified by size. A tool reporting one blended 95% sensitivity figure can be hiding near-perfect performance on large lesions and mediocre performance on the sub-5mm band that dominates incidental findings. That is why aggregate sensitivity is a observed-pattern-class comfort metric, not a clinical one: in the reading rooms and validation cohorts we have seen work through this, the size-stratified breakdown is consistently what separates a demo from a deployable tool.

Measuring it honestly requires enough small lesions in the validation set to produce a stable estimate — which is itself a data problem, because small aneurysms are under-represented in most single-site collections. This is the same generalisability challenge that federated medical benchmarking addresses; the reasoning behind it is developed in MedPerf and federated benchmarking for medical AI generalisability.

How do false positives per scan affect radiologist adoption?

This is the metric that decides whether a tool gets used at all. A detector can hit any sensitivity target you want if you accept enough false positives — recall and precision trade against each other, and pushing the operating point toward recall is easy. The problem is that every false mark costs the radiologist a dismissal, and a tool that raises several false positives per case trains readers to ignore it.

The threshold that matters is roughly one false positive per scan or fewer. A common failure pattern:

Design goal	What it optimises	What the radiologist experiences
Accuracy-first (“hit 95% sensitivity”)	Aggregate recall on study cohort	Several false marks per case; alert fatigue; tool ignored
Adoption-first (FP-per-scan budget)	Precision at a fixed sensitivity floor	Under one false mark per case; marks are trusted and reviewed

A tool that cuts false positives per case from several to under one materially changes adoption — not because the underlying model is smarter, but because the operating point was chosen against the workflow rather than against a leaderboard. This is an observed-pattern-class claim drawn from how these systems land in practice, not a published benchmark. The uncomfortable implication is that the model with the higher headline sensitivity is often the worse clinical tool.

What validation evidence proves an aneurysm-detection tool generalises?

Single-site validation is the trap. A model trained and tested on one hospital’s scanners, protocols, and patient population will report numbers that do not transfer, because scanner vendor, field strength, reconstruction kernel, contrast timing, and population demographics all shift the input distribution.

The evidence a clinical-grade tool needs is multi-site validation that spans:

• Multiple scanner vendors and models (GE, Siemens, Philips behave differently at the voxel level)
• Different acquisition protocols and reconstruction settings
• Distinct patient populations, including age and vascular-anatomy variation
• Enough sub-5mm lesions across all of the above to keep the size-stratified estimate stable

The reason to fix the operating point against a multi-site validation set early is economic as much as clinical. Programmes that design for accuracy first and confront generalisability later routinely burn a 6–12 month rework loop rebuilding evidence they could have gathered up front (a observed-pattern from medical-CV engagements, not a fixed schedule). The parent constraints — how FDA Software-as-a-Medical-Device rules turn generalisability into concrete acceptance criteria — are framed in our broader work on computer vision engineering, which this aneurysm-detection case instantiates.

How does aneurysm-detection inference integrate with PACS and the reading workflow?

The best model is worthless if it does not reach the radiologist inside their existing workflow. Integration is through the picture archiving and communication system (PACS) and the DICOM plumbing that moves studies around a hospital. A CADe result is typically returned as a DICOM secondary-capture or structured-report object that the reading workstation surfaces alongside the original study — marks the radiologist can accept or dismiss without leaving their tool.

Two engineering realities dominate here. First, latency: the result must arrive before the radiologist reads the study, which sets a wall-clock budget on the whole 3D pipeline. Second, the marks must be reviewable, not authoritative — the system flags candidates, the radiologist decides. A tool that inserts itself between the reader and their judgment fails both clinically and, under SaMD framing, regulatorily.

Under FDA SaMD rules, why is the shipped operating point a locked commitment?

Here is where aneurysm detection stops being a machine-learning problem and becomes a regulatory one. In research, the operating point — the confidence threshold that trades sensitivity against false positives — is a slider you tune freely. Once a tool is cleared as Software as a Medical Device, that slider is welded in place.

The operating point you validate is the operating point you commit to. Your clearance rests on evidence gathered at a specific threshold; the sensitivity and false-positive numbers in your submission are true only at that point on the curve. Silently nudging the threshold post-clearance changes the device’s clinical behaviour and invalidates the evidence — it is a new device, not a settings change. This lock-and-key relationship between the validated operating point and the shipped one is why accuracy-first development is not just inefficient but structurally misaligned with how these tools are regulated. The same discipline underpins how detection accuracy maps to FDA validation evidence more generally.

FAQ

What’s worth understanding about aneurysm detection first?

Aneurysm detection is computer-aided detection (CADe) on 3D CTA or MRA volumes: the system flags candidate outpouchings on cerebral arteries so a radiologist reviews them, without replacing clinical judgment. In practice it usually segments the vascular tree first and then reasons about candidates in vessel context, rather than running a raw 3D bounding-box detector.

Which CV patterns power aneurysm detection — 3D segmentation, detection, or a combined CADe pipeline?

All three, combined as a pipeline: 3D vessel segmentation (often 3D U-Net variants), high-recall candidate detection, and a separate false-positive-reduction stage. The pipeline structure is chosen over a monolithic end-to-end detector because it lets engineers isolate which stage failed when a lesion is missed or a scan is flooded with false marks.

Why is small-lesion (sub-5mm) sensitivity the hardest part of aneurysm detection, and how is it measured?

Small aneurysms are hard to detect but common as incidental findings, so aggregate sensitivity can hide poor sub-5mm performance behind strong large-lesion numbers. The honest measure is per-lesion sensitivity stratified by size, which requires a validation set with enough small lesions to produce a stable estimate.

How do false positives per scan affect radiologist adoption, and what operating point trade-off matters?

Every false mark costs a radiologist a dismissal, so a tool raising several false positives per case trains readers to ignore it. The trade-off is recall against precision: choosing the operating point against a false-positive budget (roughly one per scan or fewer) at a fixed sensitivity floor drives adoption far more than chasing a higher headline sensitivity.

What validation evidence does an aneurysm-detection tool need to generalise across scanners, protocols, and patient populations?

Multi-site validation spanning multiple scanner vendors, acquisition protocols, and patient populations — with enough sub-5mm lesions across all of them to keep the size-stratified estimate stable. Single-site numbers do not transfer, and fixing the operating point against a multi-site set early avoids a 6–12 month rework loop.

How does aneurysm-detection inference integrate with PACS and the radiology reading workflow?

Results are returned through PACS as DICOM secondary-capture or structured-report objects that the reading workstation surfaces alongside the study, as marks the radiologist can accept or dismiss. The whole 3D pipeline has a wall-clock latency budget because the result must arrive before the radiologist reads the study.

Under FDA SaMD rules, why is the shipped operating point a locked regulatory commitment rather than a tunable setting?

Clearance rests on evidence gathered at a specific confidence threshold, so the sensitivity and false-positive figures in the submission are only true at that point on the curve. Silently changing the threshold post-clearance alters clinical behaviour and invalidates the evidence — it is a new device, not a settings change.

Where this leaves the design decision

The question that decides an aneurysm-detection programme is not “how sensitive is the model?” but “at which single, locked operating point can we defend size-stratified sensitivity, a sub-one-per-scan false-positive rate, and evidence that holds across scanners we do not own?” A team that can answer that with multi-site data in hand has a clearable tool. A team optimising aggregate sensitivity has a demo — and a rework loop waiting for it. The medical-device CV regulatory pathway is where those acceptance criteria get pinned down before the first line of the submission is written.