The gains claimed for automatic assistance span two orders of magnitude, and that spread is more instructive than any single figure.On one side, a study with three physicians of ten years’ experience, comparing annotation from scratch against correcting automatic proposals, concludes that assistance attains higher annotation quality and boosts annotation speed by approximately 25 per cent, specifically saving around 1.31 minutes per image and 0.2 minutes per object. On the other, commercial sources routinely claim factors above eight.The same technology, measurements differing by more than thirty times. The difference lies not in the tool but in what is measured, by whom, and on which structures. This article covers the available methods, their real gains and their hidden costs. It extends the complete guide to medical image segmentation.
Why the figures diverge
Four factors explain most of the spread, and identifying them lets any claimed gain be read properly.The first is the comparison point. Comparing assisted correction against full manual slice-by-slice contouring gives a high ratio. Comparing it against realistic practice, where the annotator already uses interpolation and propagation from the previous slice, gives a far more modest one. It is the second comparison that reflects production.The second is the structure. A large well contrasted organ is almost entirely handled by assistance; a lesion with a gradual boundary is barely handled at all. A gain measured on the former does not transfer to the latter.The third is what the stopwatch covers. Contouring time is only part of annotation time, which includes navigation, verification, correction and validation. Halving the drawing alone yields far less than half on the total.The fourth is the operator. A fast expert gains less than an average annotator, since their baseline time is already short. Studies run with experienced clinicians therefore mechanically produce lower gains, and they are the most rigorous ones.Families of semi-automatic methods
Five approaches coexist in medical image segmentation, and a project generally combines several.Inter-slice interpolation
The annotator contours one slice in several and the tool computes the intermediates by geometric interpolation. It is the oldest method, the most predictable and often the most profitable on large structures with regular geometry.Its limit is that it ignores the image: it interpolates a shape without checking the result matches the signal. It therefore fails where geometry changes fast, at organ poles and bifurcations. The practical rule is densifying the contoured slices there rather than spacing them uniformly.Thresholding and region growing
These methods exploit intensity: the annotator indicates a starting point and a criterion, and the tool aggregates connected voxels satisfying it.They are extremely effective on homogeneous well contrasted structures, bone on CT or aerated lung, where they produce in seconds what would otherwise take hours. They become unmanageable as soon as the boundary is ambiguous, propagation leaking into neighbouring tissue of similar density, and correcting a leak then costs more than the original drawing.Active contours
An initial curve evolves under forces attracted by image gradients and constrained by smoothness. The result is a smooth contour following marked boundaries.This family suits compact structures with continuous boundaries and behaves poorly on branching objects or interrupted boundaries. Its parameterisation is also less intuitive than the others, making it less suited to a large team where setting consistency matters.Interactive click-based assistance
The annotator indicates points belonging to the structure and points not belonging to it, and a model produces a segmentation refined at each interaction. This is currently the most versatile family, because it presupposes neither homogeneity nor a sharp boundary.Its quality depends heavily on the underlying model and its fit to the modality and structure. A study on ultrasound annotation with six practitioners reports an average annotation time reduction of 34.1 per cent for liver mass segmentation.Model pre-annotation
This is the regime that has advanced most in recent years and whose landscape moves fastest. A model processes the corpus upfront and the annotator corrects. It is the most effective regime where the model fits well, and the riskiest for anchoring.Its implementation is also the simplest of the five, since it requires no interaction during the work. On structures covered by available multi-organ models, it transforms a project’s economics.The gain and its conditions
The study cited in the introduction adds a qualification worth more than a headline figure.Its authors observe that the greater the number of anatomical structures that need to be labelled, the more obvious the efficiency advantage becomes.That result is explicable. The fixed cost of an image, navigation, windowing, familiarisation with the case, does not fall with assistance; only drawing time falls. On an image with one structure, that fixed cost dominates and the relative gain stays small. On an image with ten, drawing dominates and the gain approaches the gain on drawing alone.The costing consequence is direct. A project handling a single structure per study should not expect the gains reported on multi-organ corpora, and the two situations must be estimated separately.A second condition deserves stating: the gain assumes correction is genuinely faster than drawing. On a poor proposal, correcting costs more than starting over, and an annotator without the option to erase and restart loses time. An assistance arrangement must therefore always offer that escape.Anchoring, the hidden cost of assisted medical image segmentation
The time gain is measurable and published; the quality cost is much less so and it is real.The mechanism stays constant across all assisted methods. Faced with a plausible looking proposal, an annotator accepts more readily than corrects, and that tendency grows with the apparent quality of the proposal. It peaks where no limit is obvious, which is precisely where human annotation carried the most value.The effect is hard to detect because it produces not visible errors but a systematic shift. An assisted corpus can show excellent agreement between annotators while uniformly carrying the model’s bias, and agreement indicators signal nothing.Three measures contain it. Have a fraction of the corpus annotated without assistance and compare the two populations on structure size and boundary position. Compare assisted annotations against an expert reference produced without assistance, rather than against each other. And alternate regimes within one annotator to separate a method effect from a person effect.That control arrangement is the real cost of assistance, and it belongs in the costing alongside the time gain. A project adopting assistance without strengthening it trades time for a quality it no longer measures.The effect on variability
An adjacent phenomenon in assisted medical image segmentation deserves understanding because it is frequently misread.Assistance mechanically reduces variability between annotators, since they all start from the same proposal. That reduction is real, measurable, and regularly presented as a quality gain.It is not necessarily one. High agreement obtained by convergence on a common suggestion does not say annotators see the same thing, but that they accepted the same proposal. The distinction is decisive: in the first case the corpus is reliable, in the second it is uniformly shifted.That ambiguity has a contractual consequence. A quality commitment based on inter-annotator agreement becomes less informative once assistance is used, and it benefits from being supplemented by a measure of conformity against an independent reference.The effect on the competence profile
One organisational consequence of assistance in medical image segmentation is rarely anticipated and changes recruitment as much as costing.Assisted work demands less dexterity and more judgement. Drawing a clean contour is a manual skill acquired through practice; deciding a proposal is wrong and where it is wrong is an evaluative skill, resting on knowledge of conventions and anatomy rather than on the hand.Two consequences follow. Training must shift from the gesture towards the criteria, which is faster to transmit but requires a richer case library, notably in examples of the typical errors of the model used. And the profile sought changes: a methodical annotator, rigorous on conventions, becomes more valuable than a fast and dexterous one.A third consequence is less pleasant. Correction work is cognitively less engaging than drawing, which accelerates fatigue on long tasks. Heavily assisted projects frequently observe faster attentional drift, which manifests precisely as more validations without examination. Alternating regimes across a working day is a simple and effective remedy.What does not automate
A counterpoint is needed, since assistance is sometimes presented as making human annotation residual.Three registers escape it entirely. Qualification, meaning deciding that an image genuinely shows what is sought, remains a clinical decision. Applying conventions, particularly those resting on arbitrary choices such as including an internal structure or setting the limit of an anatomical piece, follows from no signal. And discriminating the target from structures resembling it requires knowledge the model possesses only insofar as its training corpus contained it.A fourth register escapes it partially: delineation itself, on boundaries the image does not show. An incomplete fissure, a gradual infiltration or an articular interface are not in the signal, and assistance there produces a proposal with the appearance of information without being it.It is precisely those cases that control must concentrate on, not uniformly. An arrangement sampling at random across an assisted corpus mainly verifies cases where assistance worked well.Choosing the method by structure
A simple grid orients the method choice without successive trials, and it rests on three observable characteristics of the target structure.Homogeneity first. A structure of uniform density suits thresholding and region growing, the fastest methods when they work. A heterogeneous structure rules them out immediately.Boundary sharpness next. A boundary marked along its whole length suits active contours. A partly absent boundary defeats them and points towards click-based assistance, which tolerates interruptions.Geometric regularity last. A structure whose shape evolves gradually between slices suits interpolation. A branching structure or one with abrupt variation does not, and calls for image-guided propagation.Those three questions take minutes per structure and avoid testing methods one by one. They often lead to a combination rather than a single choice: region growing initialisation where it works, click correction on the difficult zones, interpolation on the regular portions. That combination is the regime actually used on mature projects, and it explains why gains measured in real conditions are more modest than those obtained by isolating one method on one structure.Measuring the gain on your own project
No published medical image segmentation figure transfers, for the four reasons set out at the start. Measuring on the project is therefore indispensable and it is simple.The measurement protocol takes a few days. Assemble a batch of around twenty representative studies, deliberately including difficult cases. Have the same annotators process it in both regimes, assisted and unassisted, in alternating order to neutralise learning. Time the total rather than the contouring alone. And compare results against an expert reference rather than against each other.That measurement produces four decisive pieces of information no external source supplies. The real gain on the project’s structures. Its breakdown, which almost always shows the gain concentrated on a few structures. The magnitude of anchoring, visible as a systematic offset between the two populations. And the rate of cases where assistance degrades rather than helps, indispensable for deciding where to enable it.That last point deserves emphasis: assistance does not have to be applied uniformly. Enabling it on structures where it is effective and disabling it elsewhere produces a better result than a single regime, and the measurement shows where to draw the line.The improvement loop
One particular medical image segmentation regime deserves separate treatment because it changes the nature of the project.In an active learning loop, the assistance model is periodically retrained on validated annotations, so its quality improves through the project. The gain therefore grows over time, which alters the economic profile: early batches cost a lot, later ones much less.Three precautions accompany that regime. The model version used must be logged per batch, otherwise inexplicable heterogeneity appears in the corpus. The evaluation corpus must be annotated outside that loop, or it measures the system’s consistency with itself. And anchoring grows with model quality, which requires strengthening control as the gain increases, contrary to intuition.That last remark is the most important. A project relaxing its control because assistance is improving does exactly the opposite of what the situation calls for.What this changes for costing
Three practical consequences follow from all the above for a medical image segmentation costing.The first is that a gain must be quoted per structure rather than globally. A uniform discount based on an average gain is untenable as soon as the corpus contains difficult structures, and it produces a project running at a loss on the part that matters.The second is that the strengthened control arrangement must appear opposite the gain. Presenting the saving without the control overhead gives a flattering and false costing, and the gap is paid in quality rather than money, which is worse.The third is that the gain on early batches is below steady state, while settings stabilise and the team learns to work with assistance. Quoting a constant gain from the first batch sets up disappointment.Presenting assistance to the client
One point of commercial posture deserves stating, because it durably distinguishes a provider.The temptation is to quote the highest gains the literature allows, since the client is comparing offers. That strategy has a flaw: the real gain is measured on the first batch, and the gap with the claim then appears in a context where it looks like a failure.The opposite posture is sounder. State that the gain will be measured on the project, propose the alternating measurement protocol described above, and commit to a price revised once the measurement is done. That proposal surprises and it reassures, because it describes a practice rather than a promise.It has a second merit. The measurement produces information the client needs independently of the contract: it tells them on which structures automation is progressing and where human annotation remains indispensable, which guides their own investment. A provider delivering that information positions itself as an adviser rather than an executor, which is precisely the difference this kind of project allows to be demonstrated.When assistance is not worth adopting
A method article should be able to say where its subject does not apply, and three situations make assistance a poor investment.The first is a small corpus. Setting up an assisted arrangement, validating it, measuring its gain and strengthening the control has a fixed cost that a few hundred studies do not amortise. Below that scale, working manually with a good protocol is faster overall.The second is a structure no available model covers. Where the target is unusual, a rare lesion type or a non-standard anatomical convention, general-purpose models produce proposals that are wrong in ways requiring more correction than drawing. The measurement protocol described above detects this in days, and the correct conclusion is to work unassisted rather than to persist.The third is a regulated project with a short timeline. Assistance adds a layer to document, a model version to trace and an anchoring risk to demonstrate as controlled. Where the regulatory burden is already heavy and the schedule tight, the simplification of working manually can outweigh the time gain.Naming these three costs a provider some volume and buys something better. A client told that assistance would not help on their corpus, with the measurement to support it, has received advice rather than a sales pitch, and that distinction is remembered.The most common mistakes
These failures recur often enough across assisted projects that naming them is usually enough to avoid them.- Adopting a published gain factor without checking what it compares.
- Timing contouring alone rather than the full annotation time.
- Transferring a gain measured on organs to lesions.
- Applying assistance uniformly rather than where it is effective.
- Not offering the annotator the option to erase a poor proposal.
- Reading reduced variability as proof of quality.
- Not maintaining a fraction of the corpus annotated without assistance.
- Controlling by random sampling on an assisted corpus.
- Relaxing control as model quality improves.
- Presenting the time saving without the associated control overhead.
