Pearl

Computer vision on dental radiographs is the engine, and the platform now sells more around it. Second Opinion (2D) and Second Opinion 3D (CBCT) detect findings on radiographs with neural network based computer vision, each cleared as a device.

The home page now presents six products: those two, Practice Intelligence (analytics on practice performance), Pearl Voice (ambient voice AI for clinical notes), RCM (insurance claims optimization) and Calibrate (clinical training and diagnostic standards). Some of those draw on the detection models, and some (revenue cycle work, analytics on practice management data, voice notes) carry value the detection models do not supply.

A concurrent read that never diagnoses, with a real override, and a commercial framing that pulls against it. Radiographs are analyzed automatically on capture and findings appear as overlays inside the practice's own imaging software; K250525 clears the panoramic product for concurrent or second reading by dental professionals.

The clinician can toggle detection types, remove a detection the model got wrong and add one it missed, and the product states its own limitation on bone level measurement, that image angulation affects accuracy. The pressure sits around the decision.

The home page sells the product on outcomes (37 percent more disease detected, 30 percent more treatment acceptance, $30,000 more monthly production per practice), and at group level Calibrate lets leadership define diagnostic standards and benchmark providers by pathology and location; where a clinician's variance from model output is a managed metric, declining an AI finding is no longer cost free. Both directions are real: undiagnosed caries and periodontal disease are a serious problem, and showing a patient the finding on their own radiograph improves informed consent.

The approach is characterized, and the training data and version practice are not. K250525 describes neural network based computer vision that identifies and marks regions of suspected findings, and the company describes detection across bitewing, periapical, panoramic and CBCT imaging, up to 18 findings per image, automated bone level measurement from crest to cementoenamel junction with color coded severity, and processing in about 30 seconds.

It says its models are trained on the world's largest collection of expertly annotated dental radiographs and cites more than 7 million verified cases. What is not described is the training set's size and composition by source, sensor, age and geography, annotator count or agreement, or any model update or versioning policy; the accuracy figures in circulation (the clearance studies, a 94 percent figure the company calls independently validated without naming the validator, and an 82 to 98 percent reader accuracy study) measure different things and nothing says which describes which product and cohort.

The detection models are the company's own cleared devices, and the parties around them are partly named. Pearl holds eight 510(k)s for its Second Opinion family (K210365 to K250525), so the models a practice uses are in house builds that cannot be quietly swapped. The host is named: the service runs on Amazon Web Services, with EU and UK data held by Pearl EU in AWS data centers in Dublin.

What is not published is a subprocessor list, whether Pearl Voice (ambient voice notes) or RCM (insurance claims optimization) use outside model providers and what data reaches them, and whether any pretrained or licensed component sits inside the detection models.

The evidence is regulatory studies of the right design for a reading aid, run by the company, with outcome claims the studies do not measure. The panoramic clearance (K250525) reports a standalone test on 795 panoramic radiographs against a four dentist consensus (lesion level sensitivity 99, 82 and 77 percent for impacted third molars, periapical radiolucency and caries) and a fully crossed multi reader study showing significant improvement with the aid; the summary does not give the reader count.

The company cites a study in which operator diagnostic accuracy rose from 82 to 98 percent with the tool, and states its accuracy was independently validated at 94 percent without naming the validator. The home page's headline figures are outcome claims: 37 percent more disease detected, 30 percent more treatment acceptance, 24 percent more care delivered and $30,000 more monthly production per practice, with no method, population or comparator stated; more treatment accepted is not the same as more disease correctly treated. No independent, prospective evaluation of the product in use is on the company's pages.

Security and residency are documented, and the question the company's own marketing raises is not answered. The data protection page describes AES-256 encryption with regular key rotation, TLS 1.2 in transit, HIPAA compliant logging of all data access, code scanning, and certified membership of the EU to US Data Privacy Framework with its Swiss and UK counterparts, which is checkable on the US Department of Commerce registry; EU and UK data is handled by Pearl EU and stored in AWS data centers in Dublin.

The gap is model training. The company describes its models as trained on the world's largest collection of expertly annotated dental radiographs, and onboarding for group practices includes ingesting up to 18 months of historical images; whether customer radiographs feed model improvement, and whether a practice or patient could decline, is not stated on the data protection page or the privacy policy.

Retention is stated generically, for as long as necessary for the purposes of collection, with no period for radiographs or derived data. Safety engineering, meaning guardrails and how a safety event is handled, is not published.

The role is correctly understood, and the instrument is not published. The dental practice is the covered entity; Pearl receives and processes protected health information on its behalf, which makes it a business associate. The company states that it provides business associate agreement documentation to enterprise buyers alongside its attestation report, its data protection page describes HIPAA compliant logging of all data access, and it publishes buyer facing guidance that gets the underlying law right, including that a practice submitting electronic transactions is a covered entity and that any vendor touching protected health information needs a signed agreement.

The agreement itself is available on request rather than published, and no review cadence, subcontractor flow down or breach notification timetable is stated. The privacy policy carries an effective date of November 2023 and its UK section still refers to the Privacy Shield framework, invalidated in 2020, while the data protection page correctly claims Data Privacy Framework certification: a stale reference in a live document rather than a compliance failure.

A SOC 2 Type II report is named in the company's own material, and it comes on request. Pearl names its attestation as SOC 2 Type II, with the type stated, and supplies the report with business associate documentation for vendor risk review; there is no public trust center where it can be opened without asking.

The data protection page describes AES-256 encryption with regular key rotation, TLS 1.2 in transit, HIPAA compliant logging of all data access, and code scanning for security vulnerabilities, and the Data Privacy Framework certification is checkable on the US Department of Commerce registry. ISO 13485 and EU MDR certification, which the page also references, are medical device quality and regulatory regimes, not information security attestations. The report's period and scope are not published.

Clearances are stated per product and modality, and each can be checked. FDA records list eight Pearl 510(k)s: K210365 Second Opinion (4 March 2022, 21 CFR 892.2070, MYN), K242522 Second Opinion CC (16 January 2025), K242600 Second Opinion Periapical Radiolucency Contours (11 April 2025), K243893 Second Opinion Pediatric (5 May 2025), K243230 Second Opinion BLE (9 May 2025, 892.2050, QIH), K243989 Second Opinion 3D (23 May 2025, QIH), K243234 Second Opinion CS (12 June 2025) and K250525 Second Opinion Panoramic (14 November 2025; caries, periapical radiolucency and impacted third molars on panoramic radiographs, patients 16 and over, concurrent or second read; predicates Denti.AI Detect K230144 and Second Opinion K210365).

The home page says the products are FDA cleared to support 2D and 3D diagnostic workflows and reports more than 26 regulatory clearances across more than 120 countries; the company describes itself as the first dental AI company cleared for both 2D and 3D radiologic analysis, which is its own claim. Practice Intelligence, Pearl Voice, RCM and Calibrate are not presented as devices.

The panoramic clearance names its subgroups and reports only that none degraded. The 510(k) summary for Second Opinion Panoramic (K250525, November 2025) evaluated consistency across age groups, gender, five US geographic regions and six panoramic manufacturers (Carestream, Gendex, Ray, Planmeca, Sirona and Vatech), and states that no subgroup exhibited statistically significant degradation in efficacy; it does not publish the per group figures.

Testing across six imaging manufacturers is a useful check for a radiograph model, since sensor and machine differences are a common source of drift. No standing governance framework, model update policy or third party audit of model behavior is published, and race and ethnicity are not among the groups named.

A published error rate with its method and denominator is on the record, and nothing contractual stands behind it. The 510(k) summary for Second Opinion Panoramic (K250525, decision 14 November 2025) is Pearl's own filing for a neural network based detector in the Second Opinion family: tested on 795 panoramic radiographs against the consensus of four board certified dentists, it reports lesion level sensitivity of 99 percent for impacted third molars, 82 percent for periapical radiolucency and 77 percent for caries, with lesion level wAFROC of 0.98, 0.81 and 0.72, and a fully crossed reader study showing statistically significant improvement with the aid.

It is a test set result for the cleared indication (patients 16 and over; concurrent or second read), not performance in use, and it covers the panoramic product only; the bitewing and periapical, pediatric, bone level and CBCT products each have their own clearance. The company's home page claims the product detects 37 percent more disease, with no method stated. No warranty covering model output, accuracy commitment, indemnity or remediation route is published.

Named integrations with the dental systems of record, corroborated by the partners' own documentation. In dentistry the systems of record are the practice management system and the imaging software, often from different suppliers, and this product reaches both.

The home page reports more than 40 integrations, and the named list includes 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 UK and Asia Pacific.

Detections render inside the host viewer, and practice management data is synced for Practice Intelligence, so data moves in both directions across the practice's systems. Open Dental documents the integration on its own site, describing images processed automatically or manually from its imaging module, and Carestream, Patterson and Oryx document it independently, which is evidence the connections run in production. Custom integrations are offered for systems with a public interface. The integration model is bilateral connectors rather than HL7 or FHIR, which is the norm in dental software and means portability depends on the vendor.

A named provider, a named region and a selectable jurisdiction, with isolation and derived data left open. The service is cloud native on Amazon Web Services, buyers can select US or European data jurisdiction, and the data protection page states that Pearl EU processes and stores EU and UK data in AWS data centers in Dublin, Ireland.

Locally, a bridge component connects the practice imaging system so radiographs are processed on capture and results return into the host viewer, with no new hardware; a standalone web application serves practices whose imaging software has no native path. For group practices, onboarding covers up to 50 offices a month and includes ingesting up to 18 months of historical images and syncing practice management data, a large one time transfer of existing patient imaging. Nothing is published on tenant isolation, a subprocessor list, backup and disaster recovery, or whether the European option covers derived data and model artifacts or only source images.

No price of any kind is published, and the returns are published in detail. There is no rate card, no per practice, per seat or per image basis, no term, minimum, renewal or escalator language, and no volume banding for groups; sales run through demonstration and negotiation.

What is published is the other side of the equation: the home page claims 30 percent more treatment acceptance, 24 percent more care delivered, more than 20 hours saved a week and $30,000 more monthly production per practice, across more than 29,000 practices. The marketing gives a numerator and no denominator, so the return it rests on cannot be computed without entering a sales process. The product lines are discoverable (Second Opinion, Second Opinion 3D, Practice Intelligence, Pearl Voice, RCM and Calibrate), sold to practices and multi office networks.

Tightly bounded coverage, with validation per image type and no results by setting. The scope is dental practices and multi office groups, across bitewing and periapical (Second Opinion, K210365), panoramic (K250525, patients 16 and over), CBCT (Second Opinion 3D, K243989), bone level (K243230) and a pediatric indication (K243893), each with its own clearance and validation. K250525 checked consistency across five US regions and six panoramic manufacturers. The settings (independent practices, DSOs and multi office networks) are named without results by setting, and the outcome figures the company markets come without a population.