A study on a thoracic lesion annotation tool publishes a figure that belongs in every costing discussion. Annotation time per case is reported as 23 minutes with a standard deviation of 12 minutes, computed across all raters and all ten cases.
Twenty-three minutes on average, twelve minutes of dispersion. The standard deviation exceeds half the mean, meaning one case can take twice as long as another with nothing abnormal having happened. Costing on the average is therefore a bet, and the question becomes who carries it.
This article covers the cost structure of a segmentation project, the variables governing it and the estimation method. It contains no prices, for a reason developed below: a unit price out of context has no meaning. It extends the complete guide to medical image segmentation.
Cost is not a price per image
The question of medical image segmentation cost is almost always badly framed, and reframing it is the first service to render. A price per image does not exist independently of what is annotated on it. The same CT study can take two minutes to delineate one contrasted organ or several hours to delineate fifteen structures of which three are lesions with ambiguous boundaries. Comparing two offers on price per image without specifying the task is comparing two painting quotes without stating the surface. Real cost breaks into two factors. A cadence, meaning time per unit produced, which depends entirely on the task and is measured. And an hourly rate, which depends on the provider, the competence level required and the volume ordered. The unit price is the product of the two, and it means something only relative to the task used to measure the cadence. That breakdown is not a methodological detail. It determines where negotiation applies: on the hourly rate, which is a commercial matter, and on the cadence, which is a specification matter over which the client has direct leverage.What published times say
A few orders of magnitude circulate in the literature and are worth knowing, bearing in mind that they do not transfer. On multi-organ abdominal segmentation, the authors of a reference corpus describe a task that, requiring annotators to delineate and check slice by slice, may take several hours per case. On a simpler two-dimensional task, a study reports an average of three minutes and twenty-nine seconds per element, implying roughly ten minutes per image, and an extrapolation of 214.9 hours of continuous manual labelling for the whole training set of 1,414 images. Those two figures bracket most situations encountered, and their ratio exceeds a factor of twenty. Neither is a reference for a given project, but they illustrate the one transferable rule: cost depends far more on the task than on the modality or the volume.Dispersion, the forgotten variable
The figure cited in the introduction deserves exploiting beyond its mean. A standard deviation above half the mean means the distribution of times is very spread. On a real corpus, that dispersion comes from three sources: intrinsic case difficulty, presence or absence of pathology, and image quality. Three practical consequences follow. The first is that a costing must rest on a distribution rather than a mean, and that a range is more honest than a point. The second is that corpus composition becomes a price parameter: a corpus with twenty per cent complex cases does not cost the same as one with none, and that proportion must be established before commitment. The third is that the pilot batch must deliberately include difficult cases, otherwise it measures the favourable tail of the distribution. That last error is the most frequent and the most expensive. A pilot of clean cases produces a flattering cadence, an attractive price, and a project that drifts as soon as production meets the reality of the corpus.The cost lines of a medical image segmentation project
A complete costing has six lines, of which only three usually appear in a unit price. Data preparation comes first in any serious costing: conversion, series triage, de-identification, queueing. Automatable, it is nonetheless a real load that costings often omit because it is invisible. Annotation itself is the best estimated line, since it is what the cadence study measures. Quality control comes next in the sequence, and its weight depends on the reliability level sought. It is not proportional to annotation: doubling the annotated volume does not necessarily double control, but raising the quality requirement increases it sharply. Scoping forms a fourth and largely overlooked line, comprising protocol writing, case library construction and calibration sessions. It is largely fixed, which is why a small corpus is proportionally expensive. Clinical expertise time forms a fifth line, distinct from the others because it draws on a scarce resource. It is counted in decisions rather than studies. Rework, finally, is a line every project should provision and very few actually do. A missing convention detected late requires already delivered batches to be redone.The variables governing cadence in medical image segmentation
Six variables explain most of the difference between projects, and answering them situates a case in minutes. The number of structures per study comes first and weighs more than anything else, each structure adding its own contouring and its own control. Dimensionality comes next: a thin-slice volume means hundreds of contours where an image produces one. Boundary sharpness determines the share of work that cannot be automated and radically separates a contrasted organ from an infiltrative lesion. The primitive requested also weighs: a bounding box costs a fraction of a pixel mask. The proportion of pathological cases is the fifth variable, and the most often underestimated. A corpus from a specialist department carries an abnormality density that multiplies time without the study count changing. The level of control required, finally, follows from clinical use and from whether the project is regulated.The cadence study
Since no published figure transfers, measuring on the project is the only serious basis, and it is simple to run. The principle is to have a representative test batch processed by the people who will do the work, under real production conditions, timing annotation and control separately. That batch must span the diversity of the corpus and deliberately include difficult cases. Four results come out, and none can be estimated otherwise. Cadence per structure, which almost always shows a few structures concentrating most of the time. Dispersion, which gives the range. The disagreement rate, which determines the control intensity needed and therefore the quality cost. And the volume of cases escalated to the clinician, which gives the expertise time requirement. Those four figures turn an approximate quote into a defensible commitment. They also have value to the client independently of the contract, since they show where their project will be expensive and which specifications they can act on.Billing modes and risk sharing
Three medical image segmentation billing modes coexist and they differ less in amount than in how they distribute risk. Hourly billing places the cadence risk on the client. If cases are harder than expected, cost rises. That mode suits exploratory projects, heterogeneous corpora and scoping phases where the task is not yet stable. Per-image or per-object billing places the cadence risk on the provider. The client knows their cost in advance, and the provider absorbs the gap between estimated and actual cadence. That mode suits stabilised tasks on corpora whose composition is known. A third regime, rarer, shares the risk explicitly: a unit price with a revision clause if the real corpus composition departs from the test batch. It is the most honest on a project whose corpus is not fully assembled at the time of commitment, which is frequent in medical work. The choice is not a preference but a function of how well the task is known. Offering a firm unit price on an unmeasured task is a bet, and a provider doing so has either overpriced to protect itself or underpriced and will be in difficulty.The effect of volume
Volume acts on unit cost through several mechanisms worth distinguishing, since they do not operate at the same moment. The first is amortisation of fixed costs. Scoping, training and case library construction cost roughly the same for a small corpus as for a large one, so their unit share falls with volume. That is the dominant mechanism on small projects. The second is learning. A team accelerates over the first few thousand units, then stabilises. Early batch cadence is therefore not steady state cadence, which must appear in the schedule as much as in the price. The third is organisation. Substantial volume justifies investments, automated preparation, control tooling, an assistance loop, which do not pay for themselves on a small corpus. Those three mechanisms explain why unit costs fall with volume, and they also mark the limit of that fall: it applies to fixed costs and efficiency, not to the incompressible contouring time, which remains the principal line.The cost of quality control
This line deserves separate treatment because it is most often handled as an adjustment variable. Its sizing follows from measured variability and the reliability level sought, not from a flat percentage. Sampled control suffices on high-agreement structures; double annotation with arbitration is required on ambiguous ones. Applying the same regime everywhere means paying for needless control on one side and insufficient control on the other. Two costing errors recur on this line. The first expresses it as a percentage of annotation time, when it depends on difficulty rather than volume. The second omits it under assisted regimes, when assistance requires control to be strengthened rather than relaxed, for the anchoring reasons covered elsewhere. A project whose costing does not show control separately is a project where, in case of difficulty, nobody will be able to say whether the problem lies in the work or in its verification.The cost of clinical expertise
This is the most constrained line and the one that frequently decides the feasibility of a medical image segmentation project. Clinician time is the chain’s scarcest and most expensive resource, and how it is mobilised determines total cost more than its rate does. Called upon to validate every study, it makes the project impossible beyond a few hundred cases. Called upon to define, arbitrate and validate by sample, it makes a corpus of several thousand possible. The costing consequence is that this line must be estimated in decisions rather than studies. The pilot batch supplies exactly that information by measuring the escalation rate, and it is among the most useful measurements it produces. One empirical observation deserves flagging. The study cited in the introduction reports almost identical annotation times between its expert and novice groups, 23 minutes with a standard deviation of 10 for experts and 22 minutes with a standard deviation of 14 for novices. On that tooled task, clinical experience did not translate into greater speed, which supports reserving clinician time for decisions rather than production.Comparing medical image segmentation offers
From a buyer’s standpoint, five questions make apparently incomparable proposals comparable. On exactly which task was the unit price computed, with the number of structures and the primitive. Is quality control included and under which regime, or billed separately. Is scoping, protocol and case library, included in the unit price or a distinct line. What is the billing mode and therefore who carries the cadence risk. And is a revision provided for if corpus composition departs from the test batch. Those five answers make offers comparable and they often reveal that the apparent price gap covers a scope gap. A proposal appreciably cheaper than the others has generally excluded a line the project will have to bear elsewhere.The cost of not annotating
A symmetrical calculation is rarely run in medical image segmentation and it often changes the conclusion: what does having no corpus cost. Three costs occur. The first is project delay, which has a value where a market or a regulatory opportunity is at stake. The second is the performance degradation an insufficient corpus imposes, paid in clinical results or in rejection of the device. The third, the most insidious, is the technical team building the corpus artisanally itself, whose cost appears nowhere because it is absorbed into salaries already committed. That third case deserves raising explicitly in a commercial discussion. A research engineer who annotates spends their time on a task they do less well than a trained team and cannot spend it on what they were recruited for. The real cost of that solution frequently exceeds outsourcing, without ever appearing in a comparison of quotes. That comparison is legitimate and must be run honestly, including where it concludes in favour of internalisation. On a small stable corpus and a task the team masters, doing it in house often remains the right answer.Reducing cost without degrading
Four levers genuinely act on medical image segmentation cost, and they all concern specification rather than price. The first is narrowing scope to the structures the clinical use actually consumes. A scoping exercise frequently removes a structure from the initial specification, and the saving exceeds any rate negotiation. The second is matching the precision level to the decision, covered elsewhere in this cluster. Precision above what the clinical threshold requires is paid for and delivers nothing. The third is differentiating control by measured difficulty rather than applying a uniform regime. The fourth is enabling assistance selectively, where measurement shows it helps. Those four share a dependence on measurements made at pilot. A project skipping that step forfeits the only substantial savings available and ends up negotiating an hourly rate, which is the weakest lever.Schedule, the forgotten cost dimension
One variable is almost always absent from price discussions in medical image segmentation although it is part of them: the timeline. A corpus delivered in three months and the same corpus delivered in three weeks do not cost the same, for two distinct reasons. The first is team sizing, a short deadline requiring more people to be mobilised and therefore more people trained on the same protocol, which raises the fixed cost of knowledge transfer. The second is quality, a short deadline shortening the period over which the protocol stabilises through arbitration, so conventions settle earlier and sometimes less well. That second reason is the more important and the less visible. An accelerated project produces its corpus before all the difficult cases have been encountered, and missing conventions are discovered after delivery rather than during. The practical consequence is to raise the deadline alongside the volume in a costing discussion, and to state that a significant compression is paid for either in cost or in quality. Presenting it as a trade-off rather than a constraint gives the client a decision to make rather than a requirement to impose.Why this article contains no prices
A medical image segmentation cost article that quotes no figure invites the question, and the answer is worth giving plainly. A unit price detached from its task is not information, it is noise. Published alongside a specification it does not name, it will be compared against offers covering different scopes, different control regimes and different corpus compositions. The comparison will be wrong in a direction nobody can detect, and the party that loses is whoever had specified honestly. There is a second reason. Rates move with volume, with the competence level required and with market conditions, so any figure published today misleads within a year. A method, by contrast, remains valid: measuring a cadence on a representative batch and multiplying by a rate is how the number is built regardless of what the number is. The practical implication for a buyer is to distrust any provider quoting a unit price before having seen the data and the specification. That figure was necessarily produced by assumption, and the assumptions are where the disagreement will later sit. A provider proposing to measure first is not being evasive; they are describing the only procedure that yields a number capable of holding.The most common mistakes
These failures recur often enough across corpus projects that naming them is usually enough to avoid them.- Asking for or quoting a price per image without specifying the task.
- Costing on a mean when dispersion exceeds half of it.
- Assembling a pilot batch with no difficult cases.
- Omitting data preparation, a real and invisible load.
- Expressing control as a percentage of annotation time rather than by difficulty.
- Relaxing control under an assisted regime.
- Estimating clinician time in studies rather than in decisions.
- Provisioning no rework.
- Offering a firm unit price on a task whose cadence has not been measured.
- Negotiating the hourly rate rather than the specification.
