icometrix

Automated segmentation and quantification of brain structures is the entire product, and eight separate regulatory approvals across eight neurological conditions confirm that what was assessed was the algorithm rather than the software around it.

The task is one humans genuinely cannot do at this precision. Measuring whole brain and grey matter volume, tracking annualised change, and quantifying white matter lesion burden across a longitudinal series requires voxel level segmentation that no radiologist performs by eye. Deep learning is not accelerating a manual process here; it is producing a measurement that would not otherwise exist in routine care.

The longitudinal comparison capability is where the technical depth concentrates. Comparing scans acquired months or years apart, on possibly different scanners with different acquisition parameters, and distinguishing genuine atrophy from measurement variation is substantially harder than segmenting a single study, and it is the capability the company patented and built its clinical proposition on.

Surrounding software exists and is thin by comparison. The connectivity component moves images and reports between scanners, archives and record systems, and the patient facing application supports symptom tracking. Neither carries value without the measurements.

Graded A. Reporting against an age and gender matched reference population is itself a modelling decision rather than a display choice, since the normative comparison is what converts a raw volume into a clinically interpretable finding.

Measurement delivered into the radiologist's own reporting workflow, which is the correct posture, with thresholds unpublished.

The output design keeps the clinician in charge in a specific and well judged way. Results arrive as colour coded segmentations in DICOM format, so the reader sees exactly which voxels the algorithm assigned to which structure and can disagree with the segmentation on the image itself. That is stronger reviewability than a numerical score, because the evidence for the number is displayed alongside it. Concise reports carry normative reference data, and prepopulated reporting templates drop structured findings into the radiologist's own report for editing rather than issuing anything independently.

The amyloid monitoring module is the most decision proximate and is handled appropriately, producing a severity score graded according to published clinical guidelines rather than a proprietary scale, so the output maps onto criteria clinicians already use to decide whether to interrupt therapy.

What is missing is calibration. No sensitivity, specificity, threshold or measurement variability figure appears in customer facing material. For quantification the relevant number is not a detection rate but the smallest change the system can reliably distinguish from noise, and a longitudinal product whose whole purpose is detecting change should publish that limit. Without it a clinician cannot know whether a reported annual volume loss is real.

Automation bias deserves a mention, since a precise number carries an authority a hedged human impression does not.

Ask for measurement reproducibility, the smallest detectable change, and performance across scanner types.

Outputs specified to the level of individual measures, with performance figures held in the literature rather than assembled for buyers.

The functional disclosure is precise. Specific measures are named rather than described in general terms: quantification and tracking of white matter hyperintensities on one sequence and hypointensities on another, annualised brain volume change reported separately for whole brain and grey matter, and comparison against a matched reference population. For the amyloid module the outputs are enumerated further, covering the number of affected sites, their locations, and the longest axis of the largest site combined into a graded severity score. A clinician can determine exactly what will appear on the report before purchase.

The approach is identified as deep learning, the amyloid model is described as trained on thousands of brain scans, and longitudinal comparison rests on patented technology, so the intellectual basis is at least located.

What is absent is performance in customer facing material. No accuracy, agreement, reproducibility or measurement variability figure is published, no model card exists, and no architecture is described. For quantification, test retest reproducibility and inter scanner agreement are the figures that determine clinical usefulness, and they exist in the peer reviewed literature rather than anywhere a buyer would look.

The normative reference dataset is not characterised at all, despite being the interpretive foundation of every report.

Ask for reproducibility and inter scanner agreement data, and the composition of the normative reference dataset.

The models are self developed over fifteen years, and the chain around them is undisclosed and now changing.

The favourable structural point is that the intelligence is built rather than assembled. Fifteen years of development across eight cleared modules, supported by pharmaceutical partnerships during the initial phase and validated in peer reviewed literature, is not a third party model with a clinical wrapper, and no external model provider sits in the inference path.

What is undisclosed is everything around it. No cloud or hosting provider, no sub processor register, no machine learning framework and no third party component was named.

Training provenance is the more consequential gap and is partly answerable. The amyloid module is described as trained on thousands of brain scans and the source is unstated, while pharmaceutical partnerships during early development suggest trial imaging may underpin some models, which carries its own consent and governance questions. Most importantly, the normative reference population is a dataset dependency as fundamental as any training corpus, since every clinical interpretation the product makes is relative to it, and neither its source nor its composition is published.

The 2026 acquisition introduces a new and unresolved dimension. Ownership by a major scanner manufacturer means infrastructure, data flows and model development may migrate into the parent, and nothing yet describes what changes.

Ask for training and normative dataset provenance and governance, the hosting provider, and what the acquisition changes in the processing chain.

A deep validation literature accumulated over fifteen years, with deployment evidence largely absent from public material.

The validation base is real and independently visible. Multiple peer reviewed studies support the multiple sclerosis module alone, an independent artificial intelligence registry lists the product with a substantial evidence set and rates its disclosure as comprehensive, and the methods have been published across several journals and validated at academic centres. Reproducibility has been a stated design priority from the outset, which matters more here than raw accuracy, because a longitudinal measurement is only useful if the variation it reports exceeds the variation the measurement itself introduces.

Independent assessment goes beyond publication. The United Kingdom health technology assessment body issued a Medtech Innovation Briefing on the multiple sclerosis measures in April 2022, which is an evaluation conducted by a national body rather than commissioned by the company.

What is missing is the deployment half. No customer count, site count, scan volume or named health system reference was located in public material, and no outcome study was found showing that using the measurements changes management decisions or patient results. For a product whose clinical argument is better informed treatment decisions in multiple sclerosis and dementia, that is the study a buyer would want.

The acquisition by a major imaging manufacturer in 2026 constitutes commercial validation of a kind, since diligence preceded it, and it is not clinical evidence.

Ask for deployment scale, named references, and any study linking the measurements to changed treatment decisions.

Named handling controls covering the full path, with the retention question left open by a design that requires retention.

The controls are stated specifically rather than gestured at: encrypted transfer, secure storage and controlled access, under both United States health privacy law and European data protection law. Encryption in transit is the right emphasis given the architecture, since the connectivity component's function is moving images between the hospital and the analysis service, and that transfer is the exposed leg.

What the published material does not address is what happens at rest and for how long. Longitudinal comparison is the core clinical capability, and comparing a scan against one acquired two years earlier requires the earlier scan or its derived measurements to still exist. The product therefore accumulates, by design, a patient level record of brain volume and lesion burden over time. No retention schedule, deletion process or data ownership statement was located.

That record is unusually sensitive. A stored series showing accelerating brain atrophy, or a quantified lesion burden trajectory, is a neurological prognosis in numerical form, with implications for insurance, employment and driving that outlast any single clinical episode.

Whether customer imaging contributes to model development is also unstated, and a company that has trained across eight conditions over fifteen years has consumed a great deal of imaging from somewhere.

Ask what is retained for longitudinal comparison and for how long, who can retrieve it, and whether customer scans train models.

Both major privacy regimes named explicitly, with specific handling controls attached and no contractual detail.

The statement is more useful than most because it names controls rather than only frameworks. Handling is described as compliant with United States health privacy law and European data protection law, with encrypted transfer, secure storage and controlled access. A vendor operating from Belgium into the United States has genuine exposure to both regimes, and stating both is what a buyer on either side needs.

The service delivery model is what makes the missing contractual detail matter. Analysis is described as offered as a service, and the connectivity component moves images from the hospital to be processed, which means identifiable brain imaging leaves the institution in the ordinary course of operation. That is a materially different posture from an in hospital appliance, and it makes the agreement, the retention period and the processing location the substantive questions rather than peripheral ones.

None of those is published. No business associate agreement template, execution requirement, negotiation stance or subcontractor position was located, and no data processing agreement terms are described for European customers.

The longitudinal design compounds it, since comparing scans across years requires retaining prior imaging and prior measurements, so a patient specific neurological history accumulates somewhere for the product to work at all.

Ownership change adds a further question about whether agreements now sit with the parent.

Ask for the agreement template, retention period for longitudinal comparison, processing location, and which legal entity contracts.

Specific controls named without an external information security attestation behind them.

What is published is more than a badge. Encrypted transfer, secure storage and controlled access are named as controls, and compliance is claimed against both United States health privacy law and European data protection law. Naming controls a buyer can test is more useful than naming a framework alone.

What is absent is independent verification of those controls. No information security management certification, service organisation controls report, trust centre, penetration testing statement or vulnerability disclosure policy was located.

The omission is notable given the company's profile rather than despite it. As a regulated medical device manufacturer it necessarily holds a quality management certification, and that governs design and manufacturing rather than information security, which is a distinction buyers frequently blur and this index does not. Two other imaging vendors indexed here hold and publish a named information security certification, so the credential is plainly obtainable in this category.

The architecture raises the stakes. Identifiable brain imaging is transferred out of the hospital for processing and longitudinal data persists for years, so the vendor holds a substantial repository of neurological patient data rather than transiently touching it.

One pre emptive note: restating privacy compliance cannot move this grade. Only an external information security attestation, or documentation available under agreement, will.

Ask whether an information security certification is held or in progress, and what documentation is available under agreement.

A decade of clearances across eight conditions, European marking under the stricter current pathway, and a first in category authorisation in 2026.

The first United States clearance was obtained in September 2016 and the portfolio has grown to eight regulatory approved solutions, which means individual submissions for individual indications rather than one authorisation stretched across a product line. European marking is held under the current medical device regulation rather than the superseded directive, and that distinction matters: the newer framework imposed substantially heavier clinical evidence and post market surveillance requirements, and a number of legacy products did not survive the transition.

The 2026 amyloid monitoring clearance is the standout. It authorised the first artificial intelligence software to detect, measure and grade amyloid related imaging abnormalities, and the company characterises it as the first computer aided detection and diagnosis authorisation in neuroradiology. That category sits above quantification in regulatory terms because the software is making a diagnostic characterisation rather than reporting a measurement, and the evidence bar is correspondingly higher. Timing is also commercially astute, since anti amyloid therapies create a monitoring obligation that did not previously exist.

A national health technology assessment briefing in April 2022 adds an independent evaluation from a body that assesses value rather than safety.

What is not enumerated publicly is the clearance numbers and the indication for use of each module, and it is a matter of public record.

Ask for the clearance list with indications, and how the acquisition affects the regulatory holder.

Demographic adjustment is engineered into the output, and no stratified performance is published.

The design feature deserves genuine credit because it addresses a real confound rather than describing an intention. Measurements are reported against an age and gender matched reference population, which means a brain volume is interpreted relative to what is expected for that patient's demographic rather than against an undifferentiated mean. Brain volume declines with age and differs systematically by sex, so an unadjusted number would misclassify patients at both ends of the age range. Building the adjustment into the output is a fairness relevant decision expressed in the product itself.

What that does not establish is whether the adjustment is adequate. The composition of the normative reference population is not published, and a reference cohort drawn predominantly from one ethnic group, one geography or one scanner generation would propagate its limits into every comparison the system makes. Normative data is the foundation of the entire clinical interpretation here, so its composition matters more for this product than for a detection tool.

No stratified performance by race or ethnicity was located, no fairness testing and no external algorithmic audit.

The technical bias axis is equally unaddressed and is severe for this modality. Magnetic resonance measurements vary with field strength, manufacturer, coil and sequence parameters, and volumetric quantification is notoriously sensitive to all of them. A longitudinal comparison spanning a scanner replacement is exactly where measurement drift masquerades as disease progression.

Ask for the normative population composition, and measurement agreement across scanner vendors and field strengths.

No contractual allocation is published, for measurements that increasingly drive therapy decisions.

A dedicated pass located no service level agreement, no accuracy or reproducibility warranty, no uptime commitment, no indemnity and no remediation position.

The mitigating structure is real and does not move the grade. Clearance across eight modules in two jurisdictions brings post market surveillance obligations, adverse event reporting and change control, so algorithm changes cannot be made silently and a regulator exists to escalate to.

What that does not address is the specific exposure, which is sharper here than for a detection tool. These measurements inform whether a multiple sclerosis patient escalates to a different disease modifying therapy, and in the amyloid monitoring case whether an Alzheimer's patient continues or stops treatment after a potentially serious adverse effect. A measurement that under reports lesion activity or misgrades severity feeds directly into a therapeutic decision with substantial consequences either way.

The service architecture adds an availability dimension absent from installed software. Because analysis is performed as a service with images transferred out and results returned, an outage or transfer failure means no result, and no availability commitment or historical uptime figure is published.

One pre emptive note: further clearances or publications cannot move this grade. Only contractual terms, or published reproducibility limits with an availability commitment, will.

Ask for the availability commitment, what is warranted on measurement accuracy, and the process when an error is identified in a prior report.

A named integration product spanning three system types, with output formats specified and the reporting workflow addressed directly.

The connectivity component is productised rather than bespoke and its scope is stated: it connects magnetic resonance scanners, picture archiving systems and record systems, handling secure data transfer, automated report generation and clinician dashboards. Reaching the record system as well as the imaging archive is what separates this from the imaging only integrations elsewhere in this lane, and it matters for a product whose measurements are consumed by neurologists managing disease over years rather than by radiologists reading a single study.

Output formats are specified rather than promised. Colour coded segmentations are returned in DICOM so they appear inside the study the radiologist opens, concise reports carry the normative comparison, and prepopulated reporting templates drop structured findings into the radiologist's own report.

That last element is the strongest interoperability feature in the record and is rarely offered. Most imaging artificial intelligence stops at delivering a result and leaves the radiologist to retype the findings into their report, which is where adoption quietly fails. Populating the report template addresses the actual bottleneck in the radiologist's day.

Graded A on named systems, named formats and reporting workflow integration. What is missing is thinner: no specific record system or archive vendor is named, no interface standard beyond DICOM is described, and no marketplace listing was located.

Ask which record systems are integrated in production and through what standard.

The delivery model is identifiable from the product description and is nowhere specified.

What can be established is that analysis is offered as a service, with a connectivity component transferring images from the hospital for processing and returning reports and segmentations. That establishes a hosted architecture in which identifiable brain imaging leaves the institution as a matter of routine, which is a materially different posture from the in hospital appliance model another vendor in this lane publishes plainly.

Everything specific is missing. No cloud or hosting provider, no region, no residency commitment, no tenancy model and no on premise option was located.

The geography makes this consequential rather than academic. The company operates from Belgium with United States operations, serves customers on both continents, and claims compliance with European data protection law, which imposes requirements on where personal data is processed and how transfers are handled. Whether European scans are processed in Europe and United States scans in the United States, or whether processing is consolidated, is exactly what a data protection officer will ask and is not published.

The longitudinal design deepens it, since prior imaging or derived measurements must persist somewhere for years to support comparison across time.

Acquisition by a large manufacturer introduces a further open question about whether infrastructure migrates to the parent.

Ask where processing occurs by region, whether an on premise option exists, and how retained longitudinal data is located.

A reimbursement route is named and no price of any kind is published.

The reimbursement disclosure is genuine and differs in character from the temporary arrangements other vendors in this lane rely on. The service qualifies under an established procedure code covering three dimensional post processing performed on an independent workstation with concurrent supervision. That is a permanent category code with assigned relative values rather than a temporary emerging technology code, which means the payment pathway is settled rather than provisional and the billing mechanics are already familiar to radiology departments.

The corresponding weakness is that the code is generic. It is not specific to this product or to brain quantification, so it carries no signal about this vendor's value and provides a buyer with no product specific payment figure.

No vendor pricing exists in public material. No licence fee, per scan charge, per site cost, per module rate, implementation fee, minimum commitment or contract term was located, and with eight condition specific modules the question of whether they are licensed individually or as a portfolio is substantive and unanswered.

The service delivery model raises a further unknown. Analysis is described as offered as a service, which usually implies per study charging, and nothing confirms it.

One new uncertainty arrives with the 2026 acquisition, since pricing may migrate into a larger manufacturer's commercial structure and bundling with imaging equipment becomes possible.

Ask for the charging unit, module licensing structure, and whether pricing is changing under new ownership.

Eight neurological conditions, each with its own regulatory approved module, which is the broadest evidenced clinical coverage recorded in this lane.

The range spans multiple sclerosis, dementia and Alzheimer's disease, amyloid related imaging abnormality monitoring, traumatic brain injury, stroke, epilepsy, brain tumours and Parkinson's disease. What distinguishes this from a list of served conditions is that the coverage is regulatory rather than rhetorical: a portfolio of eight approved solutions means each indication was assessed on its own terms, not that one algorithm was marketed against eight diseases.

The amyloid monitoring module illustrates why the breadth is strategically coherent rather than scattered. New anti amyloid Alzheimer's therapies require imaging surveillance for a specific adverse effect, which creates a monitoring obligation that did not exist five years ago, and a company already quantifying brain images across neurology was positioned to serve it. Coverage that follows where neurology is moving is more valuable than coverage that follows what is easy to segment.

Both modalities are addressed, since the software works on magnetic resonance and computed tomography, and both acute and chronic pathways are served, from stroke through to long term disease monitoring.

Geographic reach covers Europe and the United States with operations on both continents, and a patient facing application extends the reach beyond the radiology department.

What is not evidenced is the distribution. Nothing states which modules carry real deployment volume and which are approved but lightly used.

Ask which modules are in production use and at what volume.