Derived measurement
A boundary drawn to make something visible quietly becomes a diameter, a volume or a percentage. That number then enters a decision threshold and sets a follow-up interval — and by then nobody remembers it came from a contour rather than from a person with callipers.
Overview
Segmentation is usually presented as a visualisation feature: the software outlines the thing so you can see it. The outline is not the risk. The risk is the arithmetic that follows, because a boundary and a number are treated completely differently by everyone downstream.
Nobody transcribes a contour into a letter. Everybody transcribes a volume. Once a number exists it travels — into a report, a multidisciplinary meeting, a trial dataset, a follow-up interval — carrying none of the qualifications that were visible on the image it came from.
A measurement derived from a model-drawn boundary is a separate output with its own scope, version and method, and it is never displayed as though a human measured it.
And the sharper consequence, which has real clinical teeth: a derived measurement is never compared against a threshold whose evidence base was established by a different measurement method. Nodule follow-up intervals were built on manual callipers on axial slices. Software measuring a segmented volume is not measuring the same quantity, and putting its number against that threshold is a category error wearing three significant figures.
A number is a new output
Deriving it is an act with its own method, version and scope — not a rendering of the contour.
Method travels with it
Wherever the number goes, how it was obtained goes too. A measurement without its method cannot be compared with anything.
Never a threshold from elsewhere
A guideline cut-off belongs to the method that produced its evidence. Borrowing it is not a rounding error.
The moment it becomes a measurement
boundary · model-drawn
Everything after the number is what stops it being transcribed as though a radiologist had measured it. Strip those three lines and it is indistinguishable from one that was.
| Part | Rule |
|---|---|
| The value and unit | Bound together, as Key–value pair requires. Precision reflects the method, not the float — a volumetric method that repeats to ±1.1 mm does not report 8.437 mm. |
| Derivation | Required. "From a model-drawn boundary", stated, not implied by proximity to an overlay the reader may have dismissed. |
| Method | Which measurement was made — longest axial diameter, mean of two axes, segmented volume. These are different quantities with the same units. |
| Version | The model version and, separately, the device version — see Model version change and Labelling & UDI. |
| Reproducibility | Where known, on this protocol. Where unknown, stated as unknown — an unqualified number implies a precision nobody established. |
| Comparability | What this number may and may not be compared against. This is the line most often missing and the one that changes management. |
Thresholds belong to methods
A guideline cut-off is not a fact about biology. It is a fact about a measurement made a particular way, in the studies that produced the evidence. Move the method and the cut-off does not move with it.
| Situation | What the interface must do |
|---|---|
| Threshold established by a different method | Do not evaluate the threshold. Show the number with its method and say the comparison is not supported. Offer the measurement the guideline was built on, if the device produces it. |
| Prior measured manually, current derived | Show both with their methods, and do not compute a delta — see Trend, where a line breaks across a method change. |
| Prior and current both derived, same version | Change may be computed. Reproducibility still applies: a difference inside it is not a change. |
| Same method, different model version | Treated as a method change until shown otherwise. A revised segmentation model measures differently even when it measures the same thing. |
| Reader adjusted the boundary | The number becomes a hybrid: partly drawn, partly human. Recorded as such — see below. |
Trend & change over time asks is this change real? — is it larger than the measurement's own variability. This page asks a question that comes first: is this number the same kind of number as the last one? If the answer is no, the variability question never arises, because there is nothing legitimate to subtract.
When a reader adjusts the boundary
Editable contours are a reasonable feature and a quiet hazard: the number changes as the boundary moves, and nothing on screen usually records that it did.
- An adjusted boundary produces a new measurement, marked as reader-adjusted and attributed. It does not silently overwrite the derived one.
- Both are retained. The difference between what the model drew and what the reader accepted is the most useful post-market signal this feature produces — the same argument as Reader paradigm's un-overwritten first impression.
- No live number during dragging. A value that changes under the hand invites adjusting the boundary until the number reads better, which is measurement by intention.
- Reproducibility figures do not transfer to a hand-adjusted boundary; the stated variability was established for the automated method.
States
| State | Rendering |
|---|---|
| Derived | Value, unit, derivation, method, version, reproducibility, comparability. |
| Reader-adjusted | Marked as such and attributed; the original derived value retained beside it. |
| Reproducibility unknown | Said plainly. The number is shown without a significance claim. |
| Not comparable with the prior | Both values shown, no delta, and the reason stated where the delta would have been. |
| Boundary rejected by the reader | No measurement. A number derived from a boundary a clinician disagreed with has no standing. |
| Out of validated range | Withheld with a reason — a lesion below the size the method was validated for produces no number, not a small one. |
Who holds each rule
A derived measurement is a number with a method behind it. The number’s shape is held; the method is not.
| Rule | Held by | Where |
|---|---|---|
| Tabular numerals and bound units | Key–value | A quantity requires its unit and is typeset as a measurement. Asserted: KeyValue: a quantity cannot be rendered without its unit |
| The measurement is text, never only a label on an image | Key–value | A quantity renders as text in the accessible tree. An annotation burned into pixels is not one of this component’s cases. Asserted: KeyValue: a fact is not typeset as a measurement |
| Withheld measurements state the reason | Key–value | An out-of-scope value requires a reason and renders it. Asserted: KeyValue: outside-scope carries its reason |
| An adjusted boundary produces a new measurement, and both are retained | Event log | The re-measurement supersedes the original; the original stays. Neither can be edited. Asserted: EventLog: a superseded entry stays, and says it was superseded |
| A reader-entered adjustment is never rewritten | Numeric input | If an adjustment is typed, the field never rewrites it. If it is dragged, see the next row. Asserted: NumericInput: a refused entry stays on screen as typed |
| No live number during dragging | The implementer | There is no drag control in this package. A viewer that lets a boundary be dragged is yours, and so is the rule that the number waits until the drag ends. |
| Reproducibility figures do not transfer | The implementer | A claim about statistics, stated in prose beside the value. |
| Method and comparability are in the accessible name of the value | The implementer | The label string you pass. “Volume (ellipsoid, not comparable to 12 Aug)” and “Volume” are both expressible. |
| “Not comparable” is stated in text | The implementer | A sentence. |
| Reader-adjusted values are announced as adjusted | The implementer | A word in the label. |
| Derivation stated with every measurement | The implementer | provenance renders whatever you pass. Passing the method is the rule. |
A pattern page has no component underneath it, so a rule stated here is held by whoever builds against it unless a component already holds it. Of these 11 rules, 5 are held by a component this package ships, and 6 are the implementer’s. Five held, six not. The value can be made honest by construction. The word “comparable” cannot.
Do's and don'ts
The number carries where it came from and what it may be set against. It can be transcribed without losing its qualifications.
Three decimal places of false precision, no method, no provenance. It will be copied into a letter and read as a calliper measurement forever.
Prior 6.0 mm (calliper, 2023) · current
8.4 mm (derived, v3.1)
No change computed — different measurement methods.
Both numbers, both methods, no arithmetic. The reader decides what to make of it with everything in front of them.
Growth: +2.4 mm · crosses the 8 mm follow-up threshold
A delta across a method change, evaluated against a threshold built on the other method. Two category errors, one management decision.
Derived 8.4 mm · reader-adjusted 7.6 mm (Dr S. Iyer) — both retained
The disagreement between model and reader is preserved. That difference is the most useful surveillance data the feature produces.
Size: 7.6 mm
The adjusted value overwrote the derived one. Nobody can now tell what the model said, or that a human moved it.
No measurement. The lesion is below the 6 mm size this method was validated for.
Withheld with a reason — the same rule as Confidence disclosure. A number outside validated range is not a smaller number, it is no number.
Size: 4.1 mm (low confidence)
A hedged measurement below the validated range. The hedge is in the smallest text and the number is what gets transcribed.
Accessibility
- The measurement is text, never only a label on an image. A number readable solely by looking at a contour is unavailable to assistive technology and to a printout.
- Method and comparability are in the accessible name of the value, not in a tooltip — they are the qualifications that stop it being transcribed bare.
- Tabular numerals and bound units — see Typography.
- "Not comparable" is stated in text where the delta would be, never conveyed by a broken line alone (SC 1.4.1).
- Reader-adjusted values are announced as adjusted, with the attribution, so the provenance is not carried by styling.
- Withheld measurements state the reason rather than rendering an empty field — see Key–value pair.
- Readable at 320 px (SC 1.4.10) without truncating the method line, which is the first thing a narrow layout tends to drop.
Outcomes of use
What this contributes to, in the terms of Usability & context of use. These are attributes believed to contribute to an outcome; the outcome itself is settled by observing real use in a specified context, not by this page.
- Effectiveness — management decisions rest on comparisons that are legitimate. A threshold crossed by a method change is the failure here, and it is invisible at the point it happens.
- Efficiency — automated measurement is genuinely faster, which is why it gets adopted and why the qualifications must travel with the number rather than depending on anyone remembering them.
- Satisfaction — a radiologist who trusts the number enough to sign it. The first time one is caught disagreeing with a calliper, every derived number in the product gets re-measured by hand.
Clinical safety notes
Trace these in your risk file (ISO 14971) and usability engineering file (IEC 62366-1). Each bullet says what checks it: a component that holds the rule, a test in your verification file, deployment configuration, post-market surveillance, or a design review where no test can see it. A control checked by nothing is a sentence.
- Derivation stated with every measurement. (a test in your verification file) Mitigates: a model-derived number transcribed and relied on as a human measurement.
- Method stated. (a test in your verification file) Mitigates: two different quantities with the same units being treated as the same measurement.
- No threshold evaluated across a method change. (a test in your verification file) Mitigates: management decided by a cut-off whose evidence base used a different measurement.
- No delta across a method or version change. (a test in your verification file) Mitigates: instrument change read as disease progression — the sibling of Trend's rule.
- Precision reflects reproducibility. (a test in your verification file) Mitigates: false precision implying an accuracy nobody established.
- Reproducibility stated, including when unknown. (a test in your verification file) Mitigates: a change smaller than the method's own error acted on as real.
- Reader adjustments recorded, originals retained. (held by Event log) Mitigates: loss of the model-versus-reader signal, and a number whose provenance became untraceable.
- No live value while a boundary is dragged. (a test in your verification file) Mitigates: measurement adjusted until the number reads acceptably.
- Out-of-range measurements withheld rather than hedged. (held by Key–value pair) Mitigates: a caveated number below the validated range being transcribed without the caveat.
Related
- Attribution — the boundary before it became a number.
- Trend & change over time — whether a change is real, once the numbers are comparable at all.
- Data visualisation — why a line breaks across a method change.
- Key–value pair — value, unit, reference and provenance.
- Model version change — what a revised model does to numbers already issued.
- CarryForward — the reference application.