aetherAI

Genuinely dual: aetherSlide is an image management platform, the infrastructure layer, while aetherAI Hema and the lymph node detector are diagnostic algorithms. The 2026 FDA and IVDR clearances cover the PLATFORM, which triggers the same principle applied to Indica Labs and ModMed, since a laboratory could deploy the cleared platform and run no AI at all.

Graded B rather than C because unlike those precedents the company also builds its own substantive algorithms, notably a bone marrow differential system it describes as a world first, so the AI is not merely a layer bolted onto acquired infrastructure.

Assistive across the portfolio. The lymph node product detects, classifies, and quantifies metastases for pathologist review, and the hematology product produces differential counts a hematopathologist verifies. Bone marrow differential counting is a task where the manual alternative is a human counting hundreds of cells under a microscope, so automation here displaces tedium rather than judgment, and the pathologist retains interpretation. No published detail on confidence thresholds or when cases route back for manual review was located.

Training scale is disclosed for the hematology product, more than one million curated cells producing 15 subtype differential counts at a reported 94 percent accuracy, which is a specific and checkable claim. But that accuracy figure is not broken out by subtype, and rare cell types are precisely where differential counting is hardest and where a headline average conceals weakness.

Equivalent performance data for the lymph node detector and the platform was not located, and the founder's academic publication record in venues including Nature Communications is corporate credibility rather than product validation.

No stewardship framework or governance terms were located, no model or hosting arrangement named, and one characteristic of this product class deserves stating in full because it catches experienced buyers. Whole slide images are not simply pictures of tissue.

The formats produced by scanners routinely embed additional images alongside the tissue data, including a photograph of the slide label, and that label commonly carries the accession number and frequently the patient's name, written or printed by the originating laboratory.

A slide file therefore often contains direct identifiers even when the person handling it believes they are working with anonymous tissue, and removing them requires deliberately stripping those embedded layers rather than simply omitting a filename. That matters for every question this axis asks.

If images move to the vendor for support, troubleshooting, model development or validation, the assumption that they carry no identifiers is unsafe unless someone has confirmed the label layer was removed, and an image archive is a set of identified records rather than a de identified one by default.

Nothing published states what this vendor does with images it receives, whether customer slides contribute to model development, what retention applies, or whether de identification is performed and by which party. Ask who strips embedded label and macro images and at what point, whether the vendor ever holds customer slides and under what terms, and what retention and deletion apply to images received for support.

Adoption at named major institutions is the strongest signal: deployment of the platform across four Chang Gung Memorial Hospital branches, Taiwan's largest medical centre, plus development collaborations since 2018 with National Taiwan University Hospital and Chi Mei Medical Center, and a strategic partnership with Novartis Taiwan on myeloproliferative neoplasm diagnosis.

Independent survey work on public evidence for digital pathology AI indexes the lymph node product with minimal associated publication record. No peer reviewed validation study specific to these products was located, so this rests on institutional adoption rather than published performance.

Converted from Not Rated. No stewardship framework or data governance terms were located, and one characteristic of this product class deserves stating because it catches experienced buyers.

Whole slide images are not simply pictures of tissue. The formats produced by scanners routinely embed additional images alongside the tissue data, including a photograph of the slide label. That label commonly carries the accession number and frequently the patient's name, written or printed by the originating laboratory. A slide file therefore often contains direct identifiers even when the person handling it believes they are working with anonymous tissue, and removing them requires deliberately stripping those embedded layers rather than simply omitting a filename.

That matters for every question this axis asks. If images move to the vendor for support, troubleshooting, model development or validation, the assumption that they carry no identifiers is unsafe unless someone has confirmed the label layer was removed. It also means an image archive is a set of identified records rather than a de identified one by default.

Nothing published states what the vendor does with images it receives, whether customer slides contribute to model development, what retention applies, or whether de identification is performed and by which party.

Ask who strips embedded label and macro images and at what point, whether the vendor ever holds customer slides and under what terms, whether customer data is used to improve models, and what retention and deletion apply to any images the vendor receives for support.

Converted from Not Rated. No statement of business associate terms, availability or scope was located, and the prior note framed the position correctly: a vendor headquartered outside the United States, now holding clearance and pursuing global expansion, would need business associate terms established directly with each laboratory customer.

Two questions decide what those terms need to cover, and neither is answered publicly.

The first is whether the vendor receives protected health information at all. Diagnostic software deployed inside a laboratory's own environment, operating on images that never leave it, may put the vendor in a position closer to a software licensor than a business associate. Software delivered as a hosted service is the opposite. The company's product range spans both possibilities, including an on site automated scanning arrangement with hardware partners, so the answer differs by configuration rather than by vendor.

The second is support access, and it is the practical question for any foreign vendor regardless of where the software runs. Diagnostic software requires troubleshooting, and troubleshooting requires someone looking at the case that failed. If that someone is an engineer in the company's home jurisdiction, protected health information is being accessed from abroad, and that is a business associate matter and a cross border one simultaneously. It is also the pathway most often left unexamined, because procurement asks where data is stored rather than who can look at it.

Establish which entity contracts, whether the deployment is on site or hosted, who provides support and from where, and whether remote access to customer images is possible and how it is logged.

Upgraded from Not Rated, and the prior finding is overturned rather than refined. It recorded that no attestation was located. The company publishes its credential list on its own about page, and it includes ISO/IEC 27001 certification alongside ISO 13485, a national medical device licence, a quality management system manufacturing licence and a medical device business permit. The information security certification also appears as a dated milestone in the company's own timeline.

That is a genuine information security certification, independently assessed, and it is the distinction this index draws repeatedly in this lane: the quality standard governs medical device manufacture and says nothing about information security, while the certification here is the one that does. Holding both is the right answer for a company selling regulated diagnostic software.

Held at B rather than A for four reasons. There is no trust centre and no path to request a report. The certificate itself is listed rather than published, so scope cannot be read, and scope is the thing that matters most for a company with multiple product lines and a hardware partnership. No penetration testing or vulnerability disclosure was located. And there is no attestation of the kind United States laboratory buyers typically request, which matters given the stated global expansion.

One small point about how the list is presented. It includes a commercial business identifier among the certifications. That is a company registration number, not a credential, and mixing it into a list of assessed standards inflates the apparent set. Read such lists item by item rather than by length.

A meaningful 2026 milestone: the digital pathology platform secured both FDA clearance and IVDR certification, which is a harder combination than either alone and puts the company alongside PathAI and Indica Labs as platforms authorized on both sides of the Atlantic. Taiwan FDA clearance for aetherSlide came earlier.

The algorithm portfolio is less advanced regulatorily, with independent survey work indexing the lymph node product as CE-IVD General under the older IVDD directive rather than IVDR, the same distinction flagged against Visiopharm, Owkin Dx, and DoMore. Graded B rather than A because the strongest authorization covers the platform while the diagnostic algorithms sit at a lower regulatory tier.

Converted from Not Rated. No governance framework, monitoring commitment or subgroup performance analysis was located, and the prior note identified the specific edge correctly: models developed and validated predominantly on Taiwanese patient populations now carry clearance for markets far beyond that population. European in vitro diagnostic certification has since been added alongside the United States clearance, so the reach is wider still.

This is not a general concern about diversity. Histopathology models learn from stained tissue images, and the appearance of a slide is shaped by the laboratory that produced it: fixation time, processing protocol, stain manufacturer and lot, section thickness, scanner make and its colour profile. A model trained on slides from a small number of laboratories in one country encodes those conventions. The failure mode when it meets a laboratory with different practice is degraded performance that presents as an ordinary difficult case rather than as a system error, which is why it goes unnoticed without deliberate monitoring.

The patient population question sits on top of that, since disease prevalence, subtype distribution and presentation differ across populations, and a model calibrated on one will be miscalibrated on another.

One practical route exists here that does not for unregulated vendors. Because the platform holds clearance, a summary describing the validation datasets is on the public record. A buyer can read what sites, scanners and populations the validation actually covered rather than asking the vendor to characterise it.

Ask for post market performance monitoring at the customer's own site, and for the stain and scanner conditions under which performance was established.

Training scale is disclosed for the haematology product at more than a million curated cells producing a fifteen subtype differential count at a stated accuracy, which is a specific and checkable claim and more than most of this lane offers. The problem is the form of the figure rather than its size.

A single accuracy number across a fifteen class differential is dominated by the common classes, and rare cell types are precisely where differential counting is hardest and where their presence is diagnostically decisive: a handful of blasts in a sample changes the clinical picture entirely, and they are a small fraction of any accuracy denominator.

So a headline average can be high while performance on the categories that actually change management is unknown, and the classes contributing most to the number are the ones a technologist would have counted correctly anyway. Held at C rather than lower because the claim is at least specific enough to interrogate. Equivalent performance data for the lymph node detector and for the platform was not located, and no warranty, indemnity or remediation commitment attaches.

One further point belongs on the record: the founder's academic publication record in strong venues is corporate credibility rather than product validation, and the two are easy to conflate when a reader is looking for evidence. Ask for accuracy by cell subtype with the count in each class, and the false negative rate on blasts specifically.

The platform is the integration story, functioning as an image management system that hosts diagnostic algorithms and connects to scanning infrastructure, including a fully automated slide scanning solution built with Techman Robot and Hamamatsu Photonics that pairs a collaborative robot with scanning hardware. That automation of the physical slide handling step is unusual, since most pathology AI vendors assume slides arrive already digitized. Laboratory and image environment integration; no EHR integration claimed. No published connector list was located.

Converted from Not Rated. No hosting, region, tenancy, retention or subprocessor terms were located.

The product range makes this less of a single question than it first appears, and a buyer should establish which configuration is being quoted before anything else. The company sells a slide viewing and management platform, individual diagnostic applications, and an automated slide scanning arrangement developed with robotics and optics partners that necessarily sits physically in the laboratory. Those imply different architectures: hardware and software installed on site, software running in the laboratory's own environment, or analysis delivered as a service. Residency, tenancy and access differ completely between them, and none is described.

Two factors specific to this vendor add weight. Whole slide images are very large files, which shapes the architecture: moving them repeatedly across a network is expensive enough that on site or edge processing is often the practical design, and a buyer should ask whether that is the case rather than assuming a conventional cloud model. And the company operates from a jurisdiction whose data protection regime differs from those of the markets it is expanding into, so where analysis runs and where any retained copy sits are distinct questions from where the customer is.

Ask which deployment model applies to the specific products under consideration, whether any image or derived data leaves the laboratory in normal operation, where any hosted component runs, what is retained after analysis completes, and which subprocessors are involved in each configuration.

No pricing is published. The company has applied to list on the Taiwan Innovation Board, which if completed would bring exchange level financial disclosure and give buyers visibility into vendor stability comparable to Lunit, Aiforia, and VUNO. Backing from Quanta Computer, a major contract manufacturer, and Cathay Venture indicates capital depth. Nothing on per case, per model, or platform licensing cost was located.

Broader than most pathology vendors because it spans two distinct domains rather than one tumour type. Anatomic pathology is covered via the platform and the gastric cancer lymph node detector, while hematopathology is covered by the bone marrow differential system, a genuinely underserved area where far fewer AI products exist than in solid tumour histopathology. Biopharma enterprise services extend the customer base beyond hospitals. Coverage is strongest in Asia, with global expansion following the 2026 clearances.