Computer vision and automated inspection for medical device manufacturing, engineered around the two numbers that decide everything: what escapes, and what you scrap unnecessarily.
A vision system learns the defects present in its training data. The defects that reach a patient are the rare ones, the novel ones and the ones produced by a failure mode nobody anticipated, which are by definition underrepresented or absent from any historical set. A system reporting excellent accuracy against known defect classes has demonstrated competence at the problem you already solved, and says little about the one that will matter.
A high accuracy score can hide a critical defect detection problem. And before any of that, a vision system is an optics and fixturing project with a model attached.
Manual visual inspection is the incumbent, and almost no organization knows its actual performance. Human inspectors miss defects at a rate that varies by shift, fatigue, experience and how many units they have examined that hour, and the escape rate is unmeasured because the escapes are by definition not detected. A vision system is then evaluated against a baseline nobody established, which makes the business case difficult to write honestly and the comparison difficult to settle.
Underneath that sits the trade-off that governs every inspection system: tightening to catch more defects scraps more good product, and loosening to reduce scrap lets more escape.
The incumbent baseline is unmeasured
Human inspector escape rates are rarely quantified, so there is nothing credible to compare against.
Rare defects dominate the risk and not the data
The serious failure modes are underrepresented in any historical set, which is what makes them serious
Escapes and false rejects move in opposite directions
Every threshold choice trades one against the other, and only one of them is visible in the scrap report.
Presentation determines accuracy more than the model
Lighting, fixturing, focus, angle and surface finish decide what is detectable before any algorithm runs.
Product variants multiply the inspection problem
Different models, sizes, colours, materials and revisions need distinct criteria, and a model meeting a configuration outside its scope will still return a confident answer.
Inspection sits inside a validated process
Replacing or augmenting a quality control step is a process change rather than an equipment purchase.
Establish the human escape rate before you promise to beat it.
Seeding a known defect population through normal inspection, or re-inspecting a sample of passed product independently, produces a number for what current inspection actually catches. Most organizations have never done this and are surprised by the result in both directions. Without it, the vision system has no benchmark, the business case rests on assumption, and any comparison during validation becomes an argument. The exercise takes a few weeks and it is the foundation everything else is measured against.
Most underperforming inspection systems have a presentation problem rather than a model problem. If the defect is not reliably visible in the captured image, no amount of training will find it, and improving lighting, fixturing and optics frequently delivers more than any change to the algorithm. The order of work matters more here than in almost any other AI application.
Step 1
Establish the current inspection performance, including the human escape rate, before setting any target.
Step 2
Define the defect catalogue with quality, including severity, acceptance criteria and which defects must never escape.
Step 3
Engineer the image first, covering lighting, fixturing, optics, angle and repeatability of presentation.
Step 4
Verify the defect is visible in the captured image to a human reviewer before training on it.
Step 5
Choose the method against the problem, since rule-based measurement outperforms learning on dimensional checks.
Step 6
Build the defect library deliberately, including seeded and manufactured examples of rare failure modes.
Step 7
Set the threshold by modelling both consequences, because escape and false reject costs are not comparable.
Step 8
Design for the novel defect, with anomaly detection alongside classification rather than instead of it.
Step 9
Integrate with the line and the quality system so a rejection becomes a record rather than a diverted part.
Step 10
Define fail-safe behaviour for camera loss, poor image quality, unrecognized variants and inference failure, since the system must never silently treat an uninspected product as acceptable.
If a trained inspector cannot see the defect in your image, the model will not find it either.
The honest first test of any inspection concept is to show captured images to an experienced inspector without telling them which units are defective. Where they cannot identify the defect from the image, the problem is optical rather than computational, and training a model on those images produces a system that appears to work on the training set and fails on anything new. Teams routinely skip this because it is unglamorous, and it is the single most predictive early indicator of whether the project will succeed.
These capabilities span optics, modelling and line integration, in that order of influence on the outcome. The modelling is the most visible part and rarely the limiting one.
See
Imaging and Presentation Engineering
Lighting design, fixturing, optics selection and presentation repeatability, which determine what is detectable before any model exists.
Defect Visibility Assessment
Establishing whether each target defect is reliably visible in the captured image, which is the gate before training rather than after it.
Image Acquisition Architecture
Cameras, triggering, synchronization and throughput matched to line speed and the inspection station design.
Edge Deployment
Inference at the line with the latency the cycle time requires, including behaviour when the network or a central service is unavailable.
Detect and Decide
Defect Classification
Trained detection of known defect types with per-class performance reported rather than a single aggregate accuracy figure.
Anomaly Detection
Finding units that differ from normal without a labelled example, which is the only approach to defects you have not seen.
Dimensional and Rule-Based Inspection
Measurement, presence, absence and position checks, where deterministic methods are more accurate and more defensible than learning.
Threshold and Disposition Design
Operating point selected by modelling escape and false reject consequences together, with the decision documented rather than inherited.
Integrate and Prove
MES and QMS Connectivity
Inspection results, images and dispositions recorded against the unit, batch and device record rather than held in the vision system.
Human-in-the-Loop Review
Rejected units reviewed with the image and the reason presented, since a rejection nobody can interrogate becomes a rejection nobody trusts.
Validation Support
Performance evidence, challenge sets, versioned configuration and change control to the position your quality function defines.
Drift and Performance Monitoring
Tracking reject rate, class distribution and image characteristics, since material and tooling change and the model does not notice.
What CaliberFocus does, and does not do?
A validated model is not the same as a validated inspection process. The process includes image acquisition, decision logic, manufacturing integration, human review and failure handling, and all of it has to work
Vision inspection is not one technique. The third column is the method that actually suits each problem, and applying deep learning where a measurement rule would serve is the most common source of an unnecessarily fragile and unnecessarily hard to validate system.
| Inspection | What Is Being Checked | The Method That Suits It |
|---|---|---|
| Dimensional Verification | Measurements against specification tolerances | Rule-based measurement. Deterministic, accurate and straightforward to validate. |
| Presence and Absence | Component fitted, feature present, orientation correct | Rule-based or simple classification. Rarely needs learning. |
| Label and Print Verification | Text, codes, placement, legibility, correct variant | Character recognition and code reading, with a rule comparing against the expected value. |
| Surface Defect Detection | Scratches, contamination, discoloration, inclusions | Learned detection, since the defect space is open and variable. |
| Assembly and Seal Integrity | Correct assembly, seal formation, weld quality | Learned detection, frequently combined with dimensional checks. |
| Packaging and Tray Inspection | Count, placement, seal, foreign object | Mixed, with counting and placement rule-based and contamination learned. |
| Novel Defect Detection | Anything that does not look like known-good product | Anomaly detection, which is the only method for defects without examples. |
Label Inspection Is the Highest-Consequence Easy Win
A wrong label on a medical device is a serious event with a predictable path to a field action, and it is also one of the most tractable vision problems available: the expected content is known in advance from the work order, the comparison is deterministic, and the failure mode is unambiguous.
Vision systems degrade for physical reasons, not computational ones. The third column is how each change presents, and the pattern is that reject rate moves before anybody understands why.
| Change | What Happens Physically | How It Presents |
|---|---|---|
| Lighting Degradation | Illumination output falls gradually over months | Reject rate creeps upward with no process change to explain it. |
| Lens or Window Contamination | Dust, residue or film accumulates on the optical path | Localized false rejects, frequently in a consistent image region. |
| Fixturing Wear or Movement | Part presentation shifts by a small amount | Measurement drift and position-related failures. |
| Material or Supplier Change | Different resin, finish, colour or surface texture | A step change in reject rate dated to a lot or a delivery. |
| Tooling Wear | Parts change subtly as moulds or dies age | Gradual drift that mimics a model problem and is a process signal. |
| Ambient Light Change | Seasonal daylight, new overhead lighting, a door left open | Intermittent failures correlated with time of day or shift. |
Monitor the images, not only the results
Brightness, contrast and histogram statistics drift before accuracy does, and they identify the cause rather than the symptom.
Treat a reject rate change as a process signal first
Tooling wear and material change are more common explanations than model degradation and more actionable.
Keep a golden sample routine
Running known-good and known-defective units on a schedule is the cheapest verification that the system still works.
Record the image with the result
A rejection without its image cannot be reviewed, disputed or learned from, and will eventually be overridden by habit.
Design the operator experience deliberately
An operator who cannot see why a unit was rejected will start passing units through, and the system has then stopped functioning.
Plan for the edge to run alone
Line inspection cannot wait for a network, and behaviour during a disconnection is a design decision rather than an accident.
A rising reject rate is usually telling you about the process, not the camera.
When inspection rejects climb, the first assumption is that the vision system has degraded. More often the system is working correctly and reporting a genuine change: a new material lot, tooling approaching the end of its life, a supplier substitution, a process parameter that drifted.
Visual inspection is a quality control step in a controlled manufacturing process, documented in procedures somebody approved. Replacing or augmenting it with an automated system changes how that control operates, which makes it a process change requiring whatever assessment, qualification and documentation your quality system specifies. That determines the architecture, the evidence the system must produce and the change control around every subsequent model update.
Validation
- Treated as a change to a controlled process rather than as equipment installation, with the qualification your quality system requires
- Engineering controls capable of supporting the frameworks that apply, which may include FDA regulation, ISO 13485, GxP expectations and electronic record requirements
- Applicability, interpretation and validation strategy determined by your quality and regulatory functions rather than by us
- Challenge sets including known defects, borderline units and seeded failures, retained and rerun after every change
Change control
- Model retraining governed as a process change, since a new model is a different inspection with different performance
- Threshold adjustment governed, because moving an operating point changes both escape and false reject rates
- Lighting, fixturing and optics changes treated as changes to the inspection, which they are
- Versioned configuration across model, thresholds, imaging parameters and acceptance rules
Records
- Inspection result recorded against the unit, batch and device record rather than held in the vision system
- Images retained to the obligation, since a disposition questioned later rests on what was actually seen
- Operator overrides recorded with reason, since this is where a system quietly stops being a control
- Model version attached to every result, so a later investigation knows which system made the call
Operations
- Golden sample verification on a defined schedule, with results recorded rather than observed
- Performance monitored per defect class, since aggregate accuracy conceals a class that stopped working
- A named owner accountable for ongoing performance, distinct from whoever installed it
- Defined behaviour when the system is unavailable, since the line will keep running and somebody must decide how
Watch the operator override rate, because that is where the control quietly stops.
When false rejects are high and production pressure is real, operators begin overriding. Each individual override is defensible and the cumulative effect is an inspection system that reports excellent detection while units pass through on human discretion
An inspection system reporting ninety-nine percent defect detection has described half its behaviour. The same system may be scrapping four percent of good product to achieve it, which is a cost the quality figure does not mention and the one operations will feel first. Escape rate and false reject rate are a pair, they move in opposite directions, and reporting either alone misrepresents the system.
Escape rate against a known-defect challenge set
Measured against seeded or independently verified defects rather than inferred from production outcomes.
False reject rate on known-good product
The operational cost, the driver of operator overrides, and the number that decides the business case.
Performance per defect class
Because aggregate figures hide the class the system has never detected well and nobody noticed.
Novel defect behaviour
What the system does with a failure mode absent from training, which anomaly detection addresses and classification does not.
Borderline unit consistency
Whether marginal units are dispositioned consistently, since inconsistency at the boundary erodes confidence fastest.
Performance after any change
Model, threshold, lighting, fixture, material or tooling, since each one alters the system and only some are releases.
Retraining the model should not be the automatic response to every inspection problem
Establish first whether the cause was camera positioning, lighting, image quality, product presentation, insufficient data, incorrect labels, product variation, threshold selection or integration, because most inspection failures are physical rather than computational.
Build the challenge set before the system
Including seeded defects, manufactured failure modes, borderline units and every historical escape you have a record of.
Keep a defect library and grow it
Every escape and every novel defect added, so a future version proves it did not reintroduce the gap.
Rerun the challenge set after every change
Including changes nobody classified as a model change, such as a new lighting unit or a fixture adjustment.
Measure against the human baseline, honestly
Including the cases where the human performs better, since those define where review should remain
Inspection systems are usually reported as detection accuracy. The outcomes that matter are the pair of rates, whether operators still trust the system, and whether anybody would notice if it degraded.
| Category | What We Measure | Why It Matters |
|---|---|---|
| Prove the Inspection Was Reliable | Whether the defect is clearly defined, reliably captured under production conditions, measured by defect type, variants handled, image quality monitored, results linked to the correct unit, ambiguous cases reviewed, failure safe, decisions reconstructable and the complete process tested on the line | A high model accuracy score does not establish that the inspection system is production ready. |
| Escape Rate | Defects passing inspection, against a challenge set rather than inferred | The safety measure, and the one production outcomes cannot establish. |
| False Reject Rate | Good product scrapped or diverted | The cost measure, the driver of overrides, and the business case. |
| Performance Against the Human Baseline | Both rates compared to measured manual inspection | The only honest comparison, and it requires measuring the incumbent. |
| Operator Override Rate | Rejections overturned at the line, with reasons | Where the control quietly stops operating. |
| Per-Class Detection | Performance by defect type rather than in aggregate | Aggregate figures conceal the class that stopped working. |
Honest Expectation Setting
An assessment may reveal poor or inconsistent lighting, inadequate camera resolution, unclear defect definitions, limited examples of critical defects, high product variation, weak image traceability, excessive false rejects, missing production integration, unvalidated thresholds, inadequate human review, equipment changes without reassessment, or pilots tested only on curated images
We will establish the current human inspection baseline, assess whether target defects are visible in your captured images, review escape and false reject rates together, examine operator override behaviour, and identify which inspections suit rule-based methods rather than learning.
Bring us a specialty or clinical setting where clinicians are spending too much time creating notes. We will assess where ambient documentation fits, what must remain clinician controlled, how it should integrate with your EHR, and how to measure whether it is actually reducing burden.
One conversation with people who have run these deployments, and a written readiness view you can use with or without us.
- AI Agents and Workflow Automation
- Voice and Conversational AI
- Document AI and Intelligent Processing
- Generative AI and Enterprise Copilots
- AI Strategy and Governance
- HCC and Risk Adjustment Analytics
Security & Compliance
