A lesion detector reports mAP50 of 0.91 and the team celebrates. Then someone asks for the mAP50-95 breakdown, and it comes back at 0.68. That 23-point gap is not noise — it is the model telling you it finds roughly the right region but draws loose boundaries around it. For a consumer object detector that might be tolerable. For a medical-device CV pipeline where boundary precision drives measurement, staging, or downstream segmentation, it is a validation problem you want to see now, not during an FDA review cycle.

mAP50-95 is not a harder version of the same accuracy number. It is a different question. It asks: how good is your localisation, not just your detection? Once you understand what the metric actually averages, the gap between mAP50 and mAP50-95 stops being a mystery and becomes a diagnostic.

What should you know about mAP50-95 in practice?

The name is the definition, once you unpack it. mAP50-95 is the mean average precision averaged across ten IoU thresholds — from 0.50 to 0.95 in steps of 0.05. That is 0.50, 0.55, 0.60, and so on up to 0.95. At each threshold, you compute average precision (AP) treating a prediction as correct only if its Intersection over Union with the ground-truth box meets or exceeds that threshold. Then you average those ten AP values.

The consequence is what makes the metric useful. mAP50 accepts a prediction as a true positive if its box overlaps the ground truth by at least 50%. A detector that lands a box loosely over a lesion — capturing it plus a generous margin of surrounding tissue — can still clear the 0.50 bar comfortably. But at IoU 0.90, that same loose box fails outright, because 90% overlap demands the predicted boundary track the true boundary closely. mAP50-95 rolls both regimes into one number, so a model rewarded only for presence detection cannot hide behind a headline score.

This is the same COCO-style primary metric popularised by the COCO detection benchmark, and it is what most modern detector leaderboards report as the top-line figure. In practice, when a stakeholder says “the model scores 0.68 mAP,” they almost always mean mAP50-95 unless they specify otherwise — an ambiguity worth resolving before anyone signs off on a number.

How is mAP50-95 different from mAP50, and when does the gap matter?

The two metrics measure different things, and treating them as interchangeable is the most common misread we see. mAP50 measures whether you found the object. mAP50-95 measures whether you found it and outlined it tightly. The difference between them is a localisation-quality signal.

Here is the practical framing:

Metric	IoU threshold(s)	What a high score tells you	What it does not tell you
mAP50	0.50 only	The model reliably locates roughly the right region	Nothing about boundary tightness
mAP50-95	0.50–0.95, 10 steps	The model locates the region and draws tight boundaries	Which threshold band is dragging the score down (need the breakdown)
The gap (mAP50 − mAP50-95)	derived	How much of your “accuracy” is loose boxing	The clinical acceptability of that looseness (context-dependent)

A small gap — say a few points — means the boxes that clear 0.50 mostly stay tight through the stricter bands. A large gap of 15 to 25 points is the signature of loose bounding boxes (an observed pattern across detection projects, not a fixed rule): the model finds targets but consistently over- or under-draws them. Whether that gap matters depends entirely on the downstream use. If a clinician only needs a “look here” flag, loose boxes may be fine. If the box feeds a measurement, a growth-tracking comparison, or a segmentation seed, boundary error propagates into every downstream number.

We walk through the reading side of this comparison in more depth in mAP@50 vs mAP@50-95: reading detection metrics for localisation quality, and the mAP50 side of the pair has its own detailed treatment in [email protected] in medical CV.

How is the metric computed — what do IoU thresholds and PR curves actually do?

Three mechanisms stack on top of each other. Understanding each one separately makes the composite number legible.

Intersection over Union (IoU) is the geometric matching test. It is the area of overlap between a predicted box and a ground-truth box divided by the area of their union. IoU of 1.0 is a perfect box; IoU of 0.5 means the boxes agree on half their combined footprint. Each IoU threshold decides which predictions count as true positives before you score anything.

The precision-recall (PR) curve captures the confidence trade-off. As you lower the confidence threshold, you accept more predictions: recall rises but precision usually falls. Average precision is the area under that PR curve — a single number summarising performance across all confidence cut-offs. The confidence score a detector assigns each box is what sweeps this curve.

The averaging step is the mAP50-95 layer. You compute one AP value at each of the ten IoU thresholds, then take the mean. So the number carries information at three levels: how confident and correct predictions are (PR curve), how strictly you demand localisation (IoU threshold), and how the model behaves across the full strictness range (the average). A detector using YOLO-style architectures scores this way by default when you run its validation loop, which is why the numbers land in most training logs without anyone asking for them.

Why does localisation precision matter for medical CV?

Consider a detector trained to flag brain aneurysms or delineate an organ boundary on a scan. If it produces a box that captures the aneurysm plus a ring of surrounding vasculature, mAP50 rewards it — the overlap clears 0.50. But that loose box is clinically ambiguous. The measured diameter is wrong. The comparison against a prior scan is unreliable. If the box seeds a segmentation model, the segmentation inherits the error.

This is where the metric stops being a leaderboard score and becomes evidence about fitness for intended use. In our experience with computer-vision systems built for regulated clinical contexts, the localisation question is often the one that separates a model that demos well from one that holds up under scrutiny. The same class of concern shows up in aneurysm detection with computer vision, where flagging the right region is necessary but not sufficient — the boundary determines whether the flag is actionable.

The clinical point is simple to state and easy to overlook: detection tells you something is there; localisation tells you where it ends. For any task where the boundary drives a measurement, staging decision, or downstream model, mAP50-95 — and specifically its stricter bands — is the number that matters.

What does a large mAP50 vs mAP50-95 gap tell you, and how do you close it?

A large gap is almost always a boundary-quality problem, not a detection problem. The model has learned what to look for and roughly where, but not exactly where the object ends. Closing it is usually a data and objective question more than an architecture one.

Common levers, roughly in the order we reach for them:

• Annotation tightness. Loose or inconsistent ground-truth boxes teach the model loose boxes. If annotators drew generous margins, the model faithfully reproduces them. Re-annotation to a tight, documented labelling standard is the highest-leverage fix and the one most often skipped.
• Loss function. Regression losses that penalise IoU directly (GIoU, DIoU, CIoU) push the model toward tighter boxes than a plain L1 corner regression. Switching the box loss can recover several points at the high IoU bands.
• Resolution and small-object handling. At the stricter thresholds, a few pixels of boundary error matter. Higher input resolution, or slicing-aided inference for small targets, tightens boxes where they were failing at 0.85 and above.
• Threshold-band diagnosis first. Before changing anything, read the per-threshold AP breakdown. If AP collapses only above IoU 0.80, you have a fine-boundary problem. If it degrades from 0.55 upward, the boxes are broadly loose. The fix differs.

The order of operations matters for regulated programmes. Catching a 15–25 point gap early (observed range across detection engagements; not a benchmarked rate) avoids the expensive path — discovering the localisation weakness during an FDA review, then running a re-annotation and retraining loop that can add months to a lock-and-key versioned device programme.

How does the choice of detection metric feed into FDA validation evidence?

An FDA-cleared device is validated against its intended clinical use, not against a generic benchmark headline. That distinction reshapes how you report detection metrics. A regulator reviewing a lesion detector wants to see localisation quality tied to what the device claims to do — if the device measures, the evidence must show boundary precision adequate for measurement.

Reporting mAP50-95 alongside its per-IoU-threshold decomposition makes the model’s real localisation behaviour legible in exactly the way a reviewer expects. A single mAP50 figure invites the obvious follow-up question — how tight are those boxes? — and having the stricter-band data already in hand is the difference between answering it in the submission and answering it in a review-cycle deficiency letter.

This is one piece of a larger validation-evidence picture. The statistical framing of how you demonstrate a model change is safe belongs with A/B testing statistics for clinical CV models; assembling the whole evidence package to match a regulatory pathway is the work a medical-device CV regulatory pathway assessment scopes. The metric choice is not the deliverable — it is one input to a coherent argument that the device does what it claims.

FAQ

What does working with mAP50-95 involve in practice?

mAP50-95 is the mean average precision averaged across ten IoU thresholds from 0.50 to 0.95 in 0.05 steps. At each threshold a prediction only counts as correct if its overlap with the ground truth meets that bar, so the metric rewards tight localisation rather than mere presence detection. In practice it is the top-line “accuracy” number most modern detector leaderboards report.

How is mAP50-95 different from mAP50, and when does the gap between them matter?

mAP50 measures whether you found the object using a single lenient 0.50 IoU threshold; mAP50-95 measures whether you found it and outlined it tightly across ten thresholds up to 0.95. The gap between them signals localisation quality: a large gap means loose boxes. It matters most when the box feeds a measurement, staging decision, or downstream segmentation, where boundary error propagates.

How is the metric computed — what do IoU thresholds, precision-recall curves, and averaging actually do?

IoU decides which predictions count as true positives at a given strictness. The precision-recall curve captures the confidence trade-off, and average precision is the area under it. mAP50-95 computes one AP value at each of the ten IoU thresholds, then takes their mean — carrying information about confidence, localisation strictness, and behaviour across the full strictness range.

Why does localisation precision (not just detection) matter for medical CV like lesion or organ detection?

A loose box over a lesion or organ produces an ambiguous measurement, an unreliable comparison against prior scans, and a poor seed for downstream segmentation. Detection tells you something is there; localisation tells you where it ends. For any task where the boundary drives a clinical number, the stricter IoU bands are the ones that matter.

What does a large mAP50 vs mAP50-95 gap tell you about a model, and how do you close it?

A large gap (often 15–25 points; an observed range, not a benchmarked rate) is the signature of loose bounding boxes — the model detects but does not outline tightly. Common fixes, in order of leverage, are tightening annotation standards, switching to an IoU-aware regression loss such as GIoU/DIoU/CIoU, and increasing resolution or small-object handling. Read the per-threshold breakdown first to diagnose which band is failing.

How does the choice of detection metric feed into FDA validation evidence for a cleared device?

An FDA-cleared device is validated against its intended clinical use, so localisation quality must be tied to what the device claims to do. Reporting mAP50-95 with its per-IoU-threshold decomposition answers the reviewer’s inevitable “how tight are the boxes?” question inside the submission rather than during a deficiency cycle. It is one input to a coherent fitness-for-use argument, not the deliverable itself.

What are common mistakes when reporting mAP50-95 in a medical-device context?

The frequent errors are reporting a single mAP figure without saying which metric it is, presenting mAP50 alone when the clinical task needs boundary precision, and omitting the per-threshold breakdown that shows where localisation degrades. Each one invites a review question that is far cheaper to answer before submission than after.

The practical discipline is small: report both numbers, publish the per-threshold breakdown, and read the gap before anyone treats the headline as fitness for use. If the gap between finding a lesion and outlining it is where your evidence gets weak, that is exactly the localisation question a medical-device CV regulatory pathway assessment exists to surface — before a reviewer surfaces it for you.