Deep Bio

There is no product here other than the models, and that is true in a stricter sense than for most vendors granted this grade.

The company builds no viewer, no image management system, no laboratory information system and no scanner. It ships algorithms that are consumed inside other companies' platforms: Roche Diagnostics' digital pathology platform, Visiopharm's application centre and Techcyte's platform, the latter two separately indexed here. Strip the models out and nothing remains, not even a user interface, because the interface belongs to the host.

What the models do is clinical rather than clerical. DeepDx Prostate detects and localises malignancy on Gleason patterns and quantifies tumour to tissue ratio. DeepDx Prostate Pro classifies histological severity and generates Gleason grades and scores. Grading prostate cancer is a judgement pathologists train for years to make, it drives treatment selection directly, and inter observer variability in Gleason grading is a well documented problem in the literature. A model producing a grade is doing the diagnostic work rather than supporting the workflow around it.

The regulatory position confirms the reading rather than merely accompanying it. A class 3 in vitro diagnostic approval is granted to software whose output carries clinical consequence, and a regulator classifying these algorithms at that level is an external judgement that the models are the medical device.

Graded A, alongside RadiantGraph as the only records in this session to earn it, and this one on the stronger basis of the two.

The product is designed as an assistive layer, the regulator classified it that way, and one design detail shows genuine attention to how pathologists actually work.

The positioning is consistent and externally confirmed. Korean approval was granted for auxiliary software supporting cancer diagnosis, the company describes the tool as assisting pathologists in precise and efficient diagnosis, and hospital deployments describe it supporting tumour severity determination and confirmation of diagnoses. The pathologist signs the case. A model that generates a Gleason grade for a pathologist to confirm is a second opinion offered before the first is final, which is the strongest form of assistance and also the form most likely to anchor a reader.

The design detail is the heatmap. Output is a colour coded overlay showing where the model believes lesions are, alongside quantitative measures, so a pathologist can see what the model saw and check it against the tissue rather than accepting a number. Showing the basis for a conclusion is the oversight mechanism that matters, and it is what distinguishes this from a system that returns a grade with no working.

The adoption story reinforces the point in a way vendors rarely admit. The company found pathologists were not noticing the heatmaps and redesigned the reports around them, which is an acknowledgement that oversight only works if the reviewer actually engages with what is presented.

What holds this at B is that nothing published describes what happens when the model and the pathologist disagree, whether disagreements are logged, or how anchoring is guarded against when a grade is presented before the reader forms their own.

Graded B.

Function, performance and clinical basis are all stated precisely, with the architecture undisclosed.

The functional description is exact. The models analyse whole slide images of haematoxylin and eosin stained prostate core needle biopsies, detect and localise malignancy based on Gleason patterns, quantify tumour to tissue ratios, classify histological severity, and generate Gleason grades and scores, returning no grade where tissue falls outside the system. Every input, output and clinical framework is named, and the framework being a published international grading system rather than a proprietary score means a pathologist can evaluate the output against a standard they already use.

Performance is published with numbers rather than adjectives, and with the method for the grading figures, which were measured against a reference standard created by three pathologists. Processing time is stated at 30 seconds per core, which is the operationally relevant figure for a laboratory sizing throughput.

Training scale is quantified, thousands of slides at approval plus more than 500,000 core images added for quality control, and the approach is identified as deep learning.

What is absent is the architecture, the network design, the training methodology, the validation split, and any versioning or update cadence. That last point matters for a class 3 device where model changes require regulatory consideration, and nothing describes how a laboratory learns which version produced a given result.

Graded B: unusually specific about what the model does and how well, silent on how it is built.

Training data is quantified and its origin is not, which is an unusual combination and lands this mid band.

The quantitative disclosure is real and rare. Thousands of slides at the time of regulatory approval, more than 500,000 core images from the United States added subsequently for quality control, deep learning identified as the method, and reference standards described as created by three pathologists and drawn from multiple hospitals including a named national university hospital. Most vendors in this index describe training data with adjectives; this one uses numbers and names an institution.

What is missing is provenance. The United States images are the substantial acquisition and nothing states which institutions supplied them, under what agreement, with what de identification standard, or on what consent or research authorisation basis. For a company whose competitive position rests partly on geographic generalisation, the origin of the data delivering that generalisation is a fair question.

On the model itself the claim is implicitly in house, with no third party or pretrained component acknowledged, no framework named, and no architecture described. The class 3 approval implies design control documentation exists including software of unknown provenance assessment, which is a regulatory artefact rather than a public one.

The distribution chain is fully transparent by contrast, with all three host platforms named and two of them separately indexed here, so a laboratory can trace exactly whose model is running inside whose platform.

Graded C.

The best evidence base built in this session, and the reason is that a substantial part of it was produced by people with no stake in the result.

The independent element is an external validation study published in Modern Pathology in October 2022 with named academic authors from Korean institutions. Modern Pathology is the journal of the discipline's principal professional body, external validation is the study design that matters most because it tests a model on data it did not see during development, and the company describes the result as demonstrating uropathologist level performance. The company also reports presenting at the United States and Canadian Academy of Pathology annually since 2018 with published papers each year, which is a sustained record of exposing work to peer review rather than a single publication.

The vendor reported figures are specific and, unusually, include the method. Detection at 99 percent sensitivity and 97 percent specificity. Grading at 98.7 percent concordance in grade group classification and 96.9 percent in no grade classification, measured against a reference standard created by three pathologists, which is the correct construction for a task with known inter observer variability. Performance was validated against reference standards from multiple hospitals including a major national university hospital.

The training and quality control data is described quantitatively: thousands of slides at the time of approval and more than 500,000 core images from the United States added subsequently for quality control, which addresses geographic generalisation directly.

Deployment evidence exists at five South Korean hospitals through trial purchase ahead of full workflow integration.

Graded A.

Training data is described quantitatively, which is more than almost any record in this index offers, and its provenance is not described at all.

What is disclosed is scale and purpose. Thousands of slides were used at the time of regulatory approval, more than 500,000 core images from the United States were added subsequently for quality control, and the company states it continues to add data. Publishing a training corpus size in numbers rather than adjectives is rare and it lets a reader judge whether the model was built on enough material for the task.

What is absent is where it came from. No source institutions are named for the United States images, no de identification standard is stated, no consent or research authorisation basis is described, and no transfer mechanism is mentioned for material moving from United States institutions to a Korean company. Half a million core images is a substantial acquisition and pathology images carry identifiers in slide labels and metadata unless deliberately stripped.

The distinction between training and quality control use is stated and is worth noting, since the company describes the United States images as serving quality control rather than primary training, which is a meaningful difference and is asserted rather than evidenced.

On customer data nothing is published. Whether images a laboratory analyses through the model are retained, whether they inform continued development, and whether learning is confined to a customer's environment is unstated in either direction, and the delivery through host platforms makes the answer depend on arrangements nobody describes.

Graded C on the strength of the quantitative training disclosure, with provenance and customer data use unaddressed.

No published position was located, and the delivery model shapes both the exposure and the difficulty of assessing it.

Because the algorithms run inside other companies' platforms, the customer's contractual relationship is usually with the host rather than with Deep Bio, and protected data handling may be governed by the host's agreement. That is a legitimate architecture and it does not remove the question: a patient's whole slide image is processed by this company's model, and whether that processing occurs inside the host's environment or reaches Deep Bio infrastructure determines whether this vendor is a subprocessor in the data path. Nothing published states which.

The quality control data raises a second and sharper question. The company reports adding more than 500,000 core images from the United States for quality control purposes. Those are patient derived pathology images that crossed both an institutional boundary and a national one to reach a Korean company, and nothing published describes the source, the de identification standard applied, the consent or research authorisation basis, or the transfer mechanism. That is a substantial data acquisition and it is described in a sentence.

No business associate agreement is offered or described, no protected data handling summary exists, no retention position is stated for images analysed, and no statement addresses cross border transfer for customers outside Korea.

The Korean class 3 approval implies regulatory scrutiny of the software including data handling under that regime, which is assurance in one jurisdiction and is not a published posture for buyers elsewhere.

Graded D.

One credential is described in general terms, and the regulatory position carries more weight here than on most records.

The described credential is quality management. The company states it has demonstrated globally recognised standards in safety and quality management as a medical software manufacturer, which is the language used for a medical device quality management system certification. That is a genuine external assessment covering design controls, risk management and change control, and for software classified at the highest device risk tier it is a meaningful assurance about how the product is built and maintained. It is quality management rather than information security, and the two are related without being the same.

The regulatory position adds to it. Class 3 in vitro diagnostic approval in Korea and CE marking both entail scrutiny of the software lifecycle, and cybersecurity expectations now form part of medical device conformity assessment in both regimes, so material exists in the submissions.

What is absent is any information security disclosure aimed at a customer. No trust centre, no service organisation control or information security management certification, no penetration testing statement, no vulnerability disclosure policy, no subprocessor list and no incident notification commitment were located.

The delivery model complicates rather than excuses this. A customer's security review will largely assess the host platform, and the algorithm inside it is a component whose own posture is undocumented, which is precisely the supply chain visibility gap security teams are being asked to close.

Graded C.

The strongest regulatory record of any analysis vendor built in this session, and the firsts are substantive rather than promotional.

DeepDx Prostate was the first artificial intelligence pathology tool in Korea to receive class 3 in vitro diagnostic medical device approval from the Ministry of Food and Drug Safety, granted as auxiliary software for cancer diagnosis. Class 3 is the highest risk classification, reserved for devices whose failure carries serious clinical consequence, and it requires clinical evidence rather than a predicate comparison.

DeepDx Prostate Pro then received what the company describes as the world's first regulatory clearance for an artificial intelligence device that classifies prostate cancer severity by Gleason grade, again at class 3. That claim is checkable in kind and plausible: detection clearances existed before it, and grading is a materially harder regulatory proposition because the output is a clinical judgement rather than a flag. Being first through a regulator on a harder indication is a genuine credential.

Subsequent designations reinforce it. The product was designated an innovative product by Korean regulators in July of the following year and by the public procurement service, which are separate assessments by different bodies.

CE marking covers Europe, and the company states globally recognised standards in safety and quality management as a medical software manufacturer, which indicates a quality management certification.

The one gap is the United States, where no clearance was located despite the company sourcing substantial United States data for quality control. A prospective United States laboratory should establish current status directly.

Graded A.

One deliberate act addresses the most important subgroup question in this product's domain, and the conventional disclosures are absent.

The deliberate act is the United States quality control data. A Korean company training prostate models on Korean material faces an obvious generalisation question, because staining protocols, scanner types and patient populations differ between countries and prostate cancer incidence and presentation vary by ancestry. Adding more than 500,000 core images from the United States for quality control is a direct response to that, at real cost, and it is the single most substantive bias mitigation encountered in this session. The company states it plainly.

External validation published in a peer reviewed journal by academic authors is a second element, since independent testing on institutions' own material is how population generalisation is actually demonstrated.

What is missing is the reporting. No performance breakdown by patient ancestry, age, institution, scanner or staining protocol was located, no model card exists, and the headline figures are aggregates. That matters here because the reference standard itself is contested territory: Gleason grading has known inter observer variability, so a model matching three pathologists at 98.7 percent has matched a consensus that itself varies, and nothing examines whether agreement is uniform across grade groups. Concordance is likely to be lower at the boundaries between grade groups, which is exactly where treatment decisions change, and no breakdown is published.

Graded C on the strength of the geographic mitigation, with subgroup reporting absent.

Nothing published addresses liability directly, and the regulatory classification does more work here than on almost any record in this index.

Class 3 in vitro diagnostic approval is the substance. It is the highest risk classification, granted on clinical evidence, and it establishes both an approved indication and an accountable manufacturer under a device regulation regime that carries post market surveillance obligations, adverse event reporting duties and recall powers. A laboratory using the product within its approved indication has a defined regulatory relationship rather than only a commercial one, and a regulator can act if the device underperforms. That is a stronger position than a contractual warranty and it exists on this record where it does not on most.

The approved indication also allocates responsibility clearly. The software is auxiliary and assists diagnosis, so the pathologist remains the diagnosing clinician, and the heatmap output means their reliance is informed rather than blind.

What is unaddressed is the commercial layer and one specific division. When the algorithm reaches a laboratory through Roche Diagnostics, Visiopharm or Techcyte, three parties are involved and nothing states what the host warrants about the model, what the model developer warrants to the host, or where a laboratory takes a complaint about a grading error. The customer may have no contractual relationship with the party whose model produced the grade.

No indemnity, limitation or performance warranty was located, and the published concordance figures are marketing claims rather than committed thresholds.

Graded C.

Distribution through other companies' platforms is the interoperability story, and it is a real one because the platforms are the ones that matter.

Three integrations are named. Roche Diagnostics' Navify Digital Pathology, which reaches laboratories through one of the largest diagnostics companies in the world. Visiopharm's application centre, described as providing more than 100 diagnosis and image analysis solutions across more than 38 countries. And Techcyte's platform, in a 2025 partnership covering both prostate and frozen section capability. Visiopharm and Techcyte are both separately indexed here, so a reader can follow the integration in both directions.

A vendor whose algorithms run inside three independent platforms has demonstrated genuine portability rather than claiming it, and for a laboratory the practical consequence is that adoption does not require changing image management systems: the model arrives inside infrastructure already in place.

The five hospital Korean deployment is described as trial purchase ahead of full integration into pathology workflows, which indicates work on laboratory system integration beyond image analysis alone.

What holds this at B is that the integration surface is not documented in the company's own terms. No application programming interface is published, no interoperability standard is named, and nothing describes how results reach a report or a laboratory information system, which is the step that determines whether a Gleason grade generated by the model becomes part of the signed case or stays in a viewer.

Graded B.

Nothing published, and the delivery model makes the question harder rather than easier to answer.

The algorithms are consumed inside three named third party platforms, so where the model executes depends on how each host has integrated it. It may run inside the host's cloud, inside a customer's on premise deployment of the host platform, or by sending images to infrastructure the vendor operates. Those are materially different arrangements with different residency consequences, and nothing published states which applies to any of the three integrations.

The cross border dimension is specific to this record and unaddressed. The company is Korean, its named platform partners are Swiss, Danish and American, and its customers span Korea, Europe and beyond. A European laboratory running the algorithm through a Danish platform under European data protection law needs to know whether patient images reach a Korean company, and Korea's adequacy status under European data protection rules is the sort of detail that determines whether a deployment is lawful. None of it is described.

No hosting provider, region, residency option, subprocessor list, retention position or export term was located, and no availability commitment exists for a component sitting in a diagnostic pathway.

One consideration mitigates the practical impact. A laboratory contracts with the host platform in most cases, so residency may be governed by that agreement, and the host is the party with a published posture. That places the burden on the buyer to ask a question the algorithm vendor could have answered once, publicly.

Graded D on the absence.

Nothing is published. No price, no unit of charge, no licensing structure, no contract term and no minimum was located in vendor or third party material.

The distribution model makes the absence structurally awkward rather than merely inconvenient. The algorithms reach customers through other companies' platforms, so a laboratory does not buy from Deep Bio at all in most cases: it buys through Roche Diagnostics, Visiopharm or Techcyte. Whether the algorithm carries a separate licence fee, is bundled into the host platform's pricing, is charged per case analysed or per site, and whether the host takes a margin, is unstated everywhere. A pathology group evaluating artificial intelligence assisted prostate grading cannot determine from any published source what the model itself costs or who they would be paying.

The procurement designation is the one commercial signal available and it is not a price. Designation as an innovative product by the Korean public procurement service is a status that eases purchasing by public institutions in that market, which tells a Korean buyer something about the route and tells a buyer elsewhere nothing.

One question is specific to per case economics and unaddressed. The company states analysis in 30 seconds per core and a prostate biopsy typically yields ten to fourteen cores, so a per core or per case charge behaves very differently from a site licence at realistic volumes, and nothing indicates which applies.

Graded D.

The narrowest coverage in the pathology lane, deliberately so, and narrow enough that the grade has to reflect it.

The core product addresses one organ, one stain and one specimen type: prostate core needle biopsies stained with haematoxylin and eosin. That is not a limitation the company hides, it is the strategy, and depth within it is complete. Detection, localisation, quantification of tumour to tissue ratio, severity classification and Gleason grade and score generation together cover the whole diagnostic question a pathologist asks of a prostate biopsy. Returning a no grade result where tissue falls outside the Gleason system, rather than forcing a classification, is a detail that reflects genuine domain understanding.

Expansion exists and is early. Work on breast cancer lesion differentiation has been presented in research findings, and a frozen section capability was included in a 2025 platform partnership, so the pipeline is broadening beyond prostate.

Geographic reach is genuinely international despite the company's size. Korean class 3 approval anchors the home market, CE marking opens Europe, distribution partnerships reach European and global platforms, and a testimonial from a pathologist practising outside both markets indicates use further afield. United States regulatory status was not established in anything located, which matters because the company has invested in United States sourced data for quality control.

Graded C: exceptional depth in one indication, minimal breadth, with the trade made deliberately.