Pearl

Computer vision pathology detection on radiographs is the product. The company was founded specifically to build it, and the detection model is what a practice buys.

A concurrent read that never diagnoses, with a real override, and a commercial framing that pulls against it.

The mechanics are good. Radiographs are analysed automatically as they are captured and findings appear as overlays inside the imaging software the practice already uses. The clinician can toggle detection types on and off, remove a detection the model got wrong, and add one it missed. An independent reviewer at the Institute of Digital Dentistry singled out that editing tool as an acknowledgement that the model is fallible and that the dentist's judgement is final. The product also states its own limitation on bone level measurement, that image angulation affects accuracy. Very little in this index concedes a failure mode inside the interface itself.

What holds this to a B is the pressure around the decision rather than the decision itself.

The product is marketed to practices on case acceptance. The company publishes a 30 percent average lift and a breakdown by procedure: 39 percent for restorative replacements, 37 percent for scaling and root planing, 34 percent for crowns, 23 percent for root canals, 19 percent for fillings, 18 percent for implants. Partner channels describe surfacing unscheduled treatment so no patient opportunity is missed and building call lists prioritised by high value care. A detection aid sold on the revenue lift it produces is asking the clinician to accept its findings, not to weigh them.

The second pressure is structural. At group and DSO level the Calibrate product lets leadership define diagnostic standards, benchmark providers by pathology and location, and document calibration for accreditation. Where a clinician's variance from model output is a managed performance metric, declining an AI finding stops being cost free, whatever the interface permits.

Write it in both directions, because both are real. Undiagnosed caries and periodontal disease are a genuine problem, an inconsistent reader is a genuine problem, and showing a patient the finding on their own radiograph is a real improvement in informed consent over asking them to take a dentist's word. The same capability supports better consent and more treatment, and nothing published separates the two. An independent review notes that real world case acceptance gains run more moderate than the marketing suggests.

Ask whether clinician edits are logged, whether that log feeds provider benchmarking, and who reviews it.

Well characterised approach, independently validated, and three different accuracy figures in circulation.

What is disclosed is real. Computer vision detection across bitewing, periapical, panoramic and CBCT imaging, up to 18 findings per image, automated bone level measurement from crest to cementoenamel junction with colour coded severity, processing in about 30 seconds. The company describes training on what it calls the world's largest collection of expertly annotated dental radiographs and cites more than 7 million verified cases. Clearance evidence includes a standalone performance study and a fully crossed multi reader multi case study, which is the right design for a reader assistance claim.

The strongest transparency point is that an independent third party validated accuracy at 94 percent. Third party validation of a vendor accuracy claim is rare in this index and it is the main reason this is not a C.

The weakness is that the numbers do not agree across venues. The site claims over 95 percent accuracy in one place and 94 percent independently validated in another, and the record elsewhere cites a study in which operator accuracy rose from 82 to 98 percent. Those measure different things, and nothing published says which figure describes which measurement on which cohort.

The training set is also asserted rather than described. No size, no demographic or geographic composition, no annotator count or agreement statistic, and no model update or versioning policy was retrieved. For a product whose fairness claim rests on subgroup testing at clearance, the composition of what it learned from is the natural next question and it is not answered.

Establish which figure a vendor quote refers to before carrying it into a comparison.

The published detail here is above average and one absence is a choice rather than an oversight, which is a distinction worth making explicit because this index encounters both. On the published side: a dedicated patient data protection page, strong stated encryption at rest and in transit, role based access controls, audit logs, and certified membership of the transatlantic data privacy frameworks and their national extensions, which are verifiable on a public government registry rather than merely asserted.

European and United Kingdom data is handled by a separate regional entity and stored in a named region, so residency is settled rather than left to inference. The gap is model training, and the company's own marketing raises it. Models are described as trained on the largest annotated dental radiograph collection in the world, and onboarding for group practices includes ingesting up to eighteen months of historical images.

Those two facts together make the question unavoidable, and two differently phrased searches, one aimed directly at the legal pages, did not answer it. The legal pages were found, so this is not a failure to look. They are silent on it, and that distinction matters: an absence where no document exists is a publishing gap, while an absence inside documents that do exist and cover adjacent ground is a choice about scope. Retention is stated only generically and no sub processor list was found. Ask whether patient images train the models, whether that is severable, and what happens on termination.

The panoramic clearance was supported by a standalone performance study and a fully crossed multi reader, multi case study, with the vendor reporting stable performance across gender, geography, and imaging device subgroups. MRMC is the appropriate study design for reader assistance claims and its use here is the substantive point. A published study is cited in which operator diagnostic accuracy rose from 82 percent to 98 percent with the tool. Graded on evidence design and regulatory validation; this index does not re verify the underlying results, and the 95 percent accuracy and 37 percent more disease figures remain vendor stated.

Above average published detail, and silent on the one question its own marketing raises.

What is published is more than most. A dedicated patient data protection page, a privacy policy, AES-256 encryption at rest and TLS 1.2 or 1.3 in transit, role based access controls, audit logs, and certified membership of the EU to US Data Privacy Framework, its UK extension and the Swiss to US framework, which are verifiable on a government registry rather than asserted. European and UK data is handled by a separate Pearl EU entity and stored in a named region.

The gap is model training. The company markets its models as trained on the largest annotated dental radiograph collection in the world, and onboarding for group practices includes ingesting up to 18 months of historical images. Whether customer radiographs feed model improvement, and on what basis a practice or a patient could decline, is the question those two facts raise together. Two differently phrased searches, one aimed directly at the company's legal pages, did not answer it. The legal pages were found, so this is not a failure to look. They are silent on it.

That distinction matters. An absence where no document exists is a publishing gap. An absence inside documents that do exist and cover adjacent ground is a choice about scope.

Retention is stated generically, for as long as necessary to fulfil the purposes for which information was collected, with no specific period for radiographs or derived data. No subprocessor list was found.

Absence of a retrieved statement is not proof none exists. Ask directly whether patient images train the models, whether that is severable from the service contract, and what happens to images on termination.

The role is correctly understood and the instrument is not published.

The scoping is the opposite of a laboratory record and worth stating plainly. The dental practice is the covered entity. Pearl receives and processes protected health information on its behalf, which makes it a business associate, and a written business associate agreement is the governing instrument. The company states it provides dedicated business associate agreement documentation to enterprise buyers alongside its attestation report.

Credit where it is unusual: the company publishes buyer facing guidance that gets the underlying law right, including that a practice submitting electronic transactions is a covered entity, that protected health information covers imaging and verbal discussion and not only records, and that any vendor touching it needs a signed agreement. A vendor explaining its customer's obligations accurately is better behaviour than most.

What is missing is the route to an A in this index, which is publishing the instrument itself so a buyer can read what governs the relationship before entering it. The agreement is available on request rather than published. No statement of review cadence, subcontractor flow down or breach notification timetable was retrieved.

One precision point for a buyer. The published privacy policy carries an effective date of November 2023 and its United Kingdom section still refers to the Privacy Shield framework, which was invalidated in 2020 and superseded by the Data Privacy Framework that the same document elsewhere correctly claims certification under. That is a stale legal reference in a live document, not a compliance failure, but it is the sort of thing that should be caught in review.

The strongest security disclosure in this batch outside a company that files with the securities regulator, and it gets the terminology right.

The attestation is named as SOC 2 Type II, with the type stated. That sounds minor and is not. Across the revenue cycle lane audited alongside this record, five of six vendors either omitted the type, omitted the report entirely, or described themselves as SOC 2 certified, which is not a thing that exists, because SOC 2 produces an attestation report from a CPA firm and there is no certification body, certificate or registry. Naming it correctly is the baseline almost nobody meets.

Supporting controls are specific rather than gestural: AES-256 at rest, TLS 1.2 or 1.3 in transit, single sign on and multi factor enforcement, role based access, audit logs, cloud native on a named public cloud with selectable United States or European jurisdiction, and attestation and business associate documentation supplied for vendor risk review. Data Privacy Framework certification is independently checkable on the Department of Commerce registry.

One distinction a buyer must not blur, and it is the same one that applies to laboratory credentials elsewhere in this index. ISO 13485, MDSAP and EU MDR are medical device quality management and regulatory conformity regimes. They govern how the product is designed, manufactured and kept in a controlled state. They are not information security attestations and should not be counted as a second and third security credential alongside the first. Read against that filter the security stack here is one attestation, correctly described.

Held at A rather than higher qualified: there is no public trust centre portal, and neither ISO 27001 nor HITRUST is held. Documents come on request.

Among the most disciplined regulatory records in the index, and stated per product and per modality rather than as a blanket claim. Clearances span 2D intraoral (bitewing, periapical), extraoral panoramic, 3D CBCT via Second Opinion 3D, a pediatric indication down to age four, and AI segmentation. The company describes itself as the first dental AI company cleared for both 2D and 3D radiologic analysis and reports regulatory clearance in 120 countries.

Subgroup performance was evaluated and reported across gender, geography, and imaging device type as part of the panoramic clearance, which is a genuine fairness disclosure and rare in this index. Held back from A because no standing governance framework, model update policy, or third party AI audit was retrieved beyond the clearance documentation.

Independent third party validation exists and three different figures circulate, which pull in opposite directions on the same record. The validation is the reason this is not lower: an outside party assessed accuracy at ninety four per cent, and third party assessment of a vendor accuracy claim is rare enough in this index to count for something, because the vendor did not set the test.

The clearance evidence is also the right design for a reader assistance claim, comprising a standalone performance study and a fully crossed multi reader multi case study. The problem is that the numbers do not agree across venues. The site claims better than ninety five per cent accuracy in one place and ninety four per cent independently validated in another, and elsewhere a study is cited in which operator accuracy rose from eighty two to ninety eight per cent.

Those measure different things, standalone algorithm performance and assisted reader performance are not the same quantity, and nothing published says which figure describes which measurement on which cohort. A buyer carrying any of them into a comparison is comparing unlike things. The training set is asserted rather than described, with no size, composition, annotator count or agreement statistic, and no versioning policy. No warranty, indemnity or remediation commitment was located. Establish which figure a quote refers to before relying on it.

The deepest integration record in this batch, and corroborated outside the vendor's own site.

The axis needs translating for this setting. In dentistry there is usually no hospital electronic health record. The systems of record are the practice management system and the imaging software, often from different suppliers, and a tool has to reach both. This one does.

More than 40 imaging and practice management systems are supported natively, named rather than implied: DEXIS, Carestream, Planmeca Romexis, Dentsply Sirona Sidexis, Apteryx XVWeb, MiPACS, SOTA Cloud, tab32, XDR, Open Dental, Dentrix including Dentrix Enterprise, Eaglesoft, Curve, Denticon, Carestack, Oryx and Dentally, with Software of Excellence EXACT for the United Kingdom and Asia Pacific. Detections render inside the host viewer rather than in a separate application, so the clinician does not change tools mid exam.

The corroboration is what lifts this above a vendor claim. Open Dental documents the integration on its own site and describes images being processed automatically or manually from within its imaging module. Carestream, Patterson and Oryx each document it independently. Third party confirmation of an integration list is rare in this index and it is what earns the A.

Beyond detection, practice management data is synced for the analytics product and custom integrations are offered for any system exposing a public interface, quoted at around 40 business days against days for a native one.

What is not evidenced is standards based interoperability. The integration model is bilateral connectors, not HL7 or FHIR, and no interface specification, uptime commitment or data export path was retrieved. Bilateral connectors are the norm in dental software and this is not a criticism of the choice, but it means portability depends on the vendor rather than on a standard.

A named provider, a named region and a selectable jurisdiction. That combination is rare enough in this index to carry the grade on its own.

The service is cloud native on Amazon Web Services. Buyers can select United States or European data jurisdiction. European and United Kingdom traffic is transacted by a separate Pearl EU legal entity and stored in Amazon data centres in Dublin, Ireland, stated explicitly rather than described as a region in Europe. Most vendors in this index name none of these things.

The local footprint is deliberately light. A bridge component connects the practice imaging system to the service so that radiographs are processed automatically on capture and results return into the host viewer. No new hardware, no imaging system replacement, configuration handled remotely by the vendor, typically in hours to days. A standalone web application exists for practices whose imaging software has no native path.

For group practices the onboarding pattern is worth understanding before signing. Batch deployment covers up to 50 offices a month and includes ingesting up to 18 months of historical images and syncing practice management data. That is a large one time transfer of existing patient imaging into vendor infrastructure, separate from ongoing operational use, and it should be scoped in the agreement rather than treated as setup.

Not published: subprocessor list, backup and disaster recovery posture, and whether the European jurisdiction option covers derived data and model artefacts or only the source images.

No published price of any kind, and an unusually specific set of published returns. The asymmetry is the finding.

Nothing is disclosed: no rate card, no per practice or per seat or per image basis, no term length, no minimum, no renewal or escalator language, no volume banding for group practices. Sales run through demo and negotiation. A third party review places the product at the mid to upper end of the dental AI market, but a reviewer's impression is not vendor disclosure.

What is published, in detail, is the other side of the equation. The company quantifies a 30 percent average lift in case acceptance and breaks it down by procedure, and its analytics product is sold on production and case acceptance reporting. A buyer is therefore given a precise numerator and no denominator, and cannot compute the return the marketing is built on without entering a sales process.

That pattern recurs across this index and is worth naming as an asymmetry a buyer can act on rather than as bad faith. The published outcome figures may well be accurate. They are simply not usable for comparison until a price sits beside them.

Ask for the pricing basis first, before the outcome deck, and ask specifically whether pricing varies with image volume, since a tool that analyses every radiograph on capture has a cost curve tied to imaging volume rather than to headcount.

Tightly and clearly bounded: dental practices, across bitewing, periapical, panoramic, and CBCT imaging, with an explicit pediatric indication from age four. Narrow scope precisely stated is what this axis rewards.