Deep Bio

The models are the product. Deep Bio sells deep learning algorithms: DeepDx Prostate for core needle biopsies, with prostatectomy, TURP and breast versions, frozen section and IHC scoring models, and DeepCDx for drug development. They reach laboratories mostly inside other companies' platforms, with integrations announced with Roche, Visiopharm, Indica Labs' HALO AP, PathAI and Tribun Health. The company's own DeepDx Viewer exists to display and manage slides for its algorithms, and it says other viewers can be used instead. Remove the models and there is nothing left to sell.

The artificial intelligence is the product.

The pathologist can correct the output, and the threshold for relying on it is not published. DeepDx Prostate overlays Gleason patterns 3, 4 and 5 in color and quantifies pattern proportions and tumor length, and its annotation tools let the user modify the AI result, with the metrics recalculated as edits are made. The company lists first read, concurrent read and quality control among its use settings; in first read use, the model's grade reaches the pathologist before their own. The regulatory approvals are for diagnostic support. Nothing states the confidence below which the software defers, how it handles tissue it cannot grade, or how a first-read grade the pathologist overrides is recorded.

Only the threshold is missing from the oversight structure.

The approach is published, with no product version or update practice. The company describes DeepDx Prostate as an artificial neural network trained on thousands of cases, and its detection and grading system is described in peer reviewed papers it lists, including Scientific Reports (2022), npj Digital Medicine and Cancers. No version history, change notice or retraining practice is published for the approved products, and nothing says how a model update is handled across the host platforms that deliver it.

Approach named without the version and update discipline.

An in-house model delivered through named host platforms, with the data path unstated. Deep Bio builds its own models, so no outside model provider sits between a slide and a grade, and it names the platforms that deliver them: Roche, Visiopharm, Indica Labs' HALO AP, PathAI and Tribun Health. Nothing says whether the model runs inside each host's environment or whether images are sent to Deep Bio, where its own DeepDx Viewer stores uploaded slides, or which subprocessors apply.

A structurally short chain at the model layer, with hosts named and the data path open.

An independent external validation is published in a peer reviewed journal. In Modern Pathology (2022), academic authors tested DeepDx Prostate on 593 prostate biopsy slides (130 benign, 463 adenocarcinoma) against grades set by three uropathologists: it detected cancer about as accurately as the original pathology reports and agreed more closely with the reference on grade group (kappa 0.713, quadratic-weighted 0.922, against 0.619 and 0.873 for the reports), and a general pathologist was tested with and without it. The company's own studies appear in Scientific Reports, npj Digital Medicine and Cancers. The external study is retrospective on one assembled dataset, so performance across scanners and laboratories rests on the developer's own work, and the company's figures of 99 percent sensitivity, 97 percent specificity and testing on 700,000 cores at US CLIA laboratories are given without a published method.

Independently evaluated performance, peer reviewed, with the method available.

General assurances only. Announcing its ISO/IEC 27001 certification, the company said it handles a great deal of medical data and has built a thorough security management system. The published privacy policy covers the website and personal details submitted through it. Nothing says whether customer slides analyzed through DeepDx are retained or used to improve models, where the 700,000 US cores it says it tested came from or how they were de-identified, or how slide label images are handled. No safety engineering is published.

General assurances that do not answer what reaches a model.

No status is stated and no privacy document reaches the product. The only privacy policy covers the website and the personal details visitors submit, and no business associate position, data processing agreement or Korean equivalent for patient images is published. In the United States the products are for research use only, and most laboratories reach them through a host platform, so whether Deep Bio ever handles protected health information depends on arrangements it does not describe.

No statement of status and no privacy document that reaches the product.

A recognized certification named, with a renewal question to raise. Deep Bio announced ISO/IEC 27001:2013 certification of its information security management system in October 2022 and said it would keep renewing it; it also holds ISO 13485 and MDSAP certification for its quality system, which are not security audits. No certificate, scope or expiry is published, and the 2013 version of the standard has since been replaced, so whether the certificate was renewed under the 2022 version is not stated. There is no trust center.

A recognized certification named without the artifact, with its clock to check.

Korean and European authorizations, with the US position stated. The company states that DeepDx Prostate is approved by Korea's Ministry of Food and Drug Safety for clinical use, where it was the first AI pathology software approved as a class 3 in vitro diagnostic, and is CE marked, and that its products are not cleared by FDA and are for research use only in the United States. It also holds MDSAP certification for its quality system. FDA's databases list no Deep Bio device. The authorizations cover the prostate product; the breast, frozen section and IHC models carry no stated clinical approval.

Non-US regimes named, with the marketing matching them.

Independent testing exists, with no results by group. The Modern Pathology validation tested the prostate model against three uropathologists on an external dataset, and the company says it has tested the product on 700,000 cores at US CLIA laboratories, which bears on performance outside the Korean material it was built on. No results are published by laboratory, scanner, stain, patient ancestry or age, and agreement is reported as an aggregate, although disagreement in Gleason grading concentrates at the boundaries between grade groups, where treatment changes. No governance framework or monitoring commitment is published.

Testing published without an evaluation by group.

The pathologist can edit the result, and nothing stands behind it. DeepDx Prostate shows its findings as Gleason pattern overlays that the pathologist can annotate and correct, and the approvals are for diagnostic support, leaving the diagnosis with the pathologist. The company's 99 percent sensitivity and 97 percent specificity are published without a denominator or method, though the full text of the independent validation may supply both. When the model reaches a laboratory through a host platform, nothing states whether Deep Bio or the host answers for a grading error. There is no published warranty, indemnity or remediation commitment.

Challenging an output is possible, and no commitment backs it.

Named image management platforms carry the output, with no laboratory or record system named. DeepDx Prostate is integrated with Roche's digital pathology platform, Visiopharm, Indica Labs' HALO AP, PathAI and Tribun Health, according to the company's announcements, so results appear inside the image systems pathologists already use. DeepDx Viewer offers reporting and AI analysis access for sites without a host. No laboratory information system integration, report format or standard is named.

Named systems with one directional delivery.

Delivered through host platforms, with location left to them. Most laboratories run DeepDx inside Roche, Visiopharm, Indica Labs, PathAI or Tribun Health environments, so hosting follows the host, and DeepDx Viewer accepts uploaded slides without saying where they are stored. No hosting provider, region or on premises option is stated for Deep Bio's own route.

Hosting implied rather than committed.

No price is published, and the sales routes can be established. DeepDx reaches most laboratories through named host platforms, Roche, Visiopharm, Indica Labs, PathAI and Tribun Health, and in Korea through purchase agreements with hospitals under the Public Procurement Service's innovation product program. DeepCDx is sold as client-driven development for biopharma. No per case, per core or license price is published, and nothing says whether a host platform bundles the algorithm or charges separately.

The sales posture is discoverable, with no price.

Coverage is named product by product, with validation behind prostate biopsy. The approved product analyzes H&E prostate core needle biopsies; prostatectomy and TURP versions are listed, along with breast resection and sentinel node models, a frozen section model across prostate, breast, stomach and lung, IHC scoring for Ki-67, ER, PR, HER2, PD-L1 and c-MET, and an ovarian BRCA prediction model. Peer reviewed validation, including an independent external study, covers prostate biopsy; the other models are presented through posters and product pages without published validation or regulatory approval.

Coverage named with validation behind part of it.