Nucs AI

Deep learning on imaging is the entire product. All three tools do inference on PSMA PET/CT: automated lesion detection and whole body tumor burden quantification, response prediction to PSMA targeted radioligand therapy, and automated longitudinal response evaluation with lesion by lesion tracking. The company reports reducing manual analysis from roughly 30 minutes to about 1 minute, which is the kind of task no rules based software performs.

Positioned as decision support with the physician retaining the decision, and the products output quantified assessments and predictions rather than directives. The company formed a Medical Advisory Board spanning radiology, nuclear medicine, oncology, and clinical research, stating explicitly that the use of AI in clinical decision making places a higher premium on clinical oversight and evidence generation. Establishing that structure before broad commercialization is the right sequence, and the investigational labeling on the prediction product reinforces that a clinician is meant to be in the loop.

Products are described functionally and the imaging modality and clinical context are specific, but no model architecture, training data provenance, or performance figures are published on the company's own materials beyond references to preliminary multicenter results. The published scientific citations relate to the underlying clinical problem rather than to the models themselves.

A data partner is named, which is more than most vendors offer and is what lifts this above the floor. A stated partnership with an imaging data aggregator is described as combining large scale imaging and clinical data with the company's models to support biomarker development and real world evidence.

Naming where training and validation data comes from converts a vague reference to proprietary datasets into a specific and answerable question: what consent or authorisation the source institutions gave, what de identification standard was applied before the data reached the model, and whether the resulting models are constrained in how they may be reused. A buyer can now ask about a named party rather than about an abstraction. The second point is the shape of the data itself.

Whole body nuclear medicine imaging is maximally identifiable material, and it is uploaded to the vendor's platform rather than processed in place, so custody genuinely transfers rather than being notional. Establish what happens to a study after the report is generated, because the institution needs the result rather than the copy retained indefinitely. One caution when reading this company's privacy surfaces belongs on the record.

The cookie consent implementation is unusually careful, granular and revocable, and it governs the marketing website: a well run website privacy control is not a statement about clinical imaging held on the platform, and the two should not be read together. Ask for retention, the de identification standard, and the training position.

Evidence is early and honestly labeled as such. The company states preliminary multicenter results show its prediction model identifies which metastatic castration resistant prostate cancer patients will respond to Lutetium PSMA therapy, and cites collaborations with Johns Hopkins and other research institutions.

The strongest external validation is commercial rather than clinical: a pharmaceutical collaboration to develop response prediction models for a PSMA targeted radioconjugate, which is a substantial third party bet on the platform. What does not yet exist publicly is a peer reviewed prospective validation of the products as deployed.

No retention period, no de identification standard and no training use position was located. Two things partially fill the picture and both should be pressed further.

The company names a data partner. A stated partnership with Segmed, an imaging data aggregator, is described as combining large scale imaging and clinical data with the company's models to support biomarker development and real world evidence. Naming where training and validation data comes from is more than most vendors offer and it is credited. It also converts a vague reference to proprietary datasets into a specific question: what consent or authorisation the source institutions gave, what de identification standard was applied before the data reached the model, and whether the resulting models are constrained in how they may be reused.

The second is the shape of the data itself. Whole body PSMA PET/CT is maximally identifiable imaging, and it is uploaded to the vendor's platform rather than processed in place, so custody genuinely transfers. Establish what happens to a study after the report is generated, since the institution needs the result rather than the copy retained indefinitely.

One caution when reading this company's privacy surfaces. The cookie consent implementation is unusually careful, granular and revocable, but it governs the marketing website. A well run website privacy control is not a statement about clinical imaging held on the platform, and the two should not be read together.

No commitment was located in public materials, and this vendor needs the axis read across three different relationships rather than one, because its stated customer base spans clinicians, researchers, device manufacturers and pharmaceutical partners.

For clinical sites uploading studies, ordinary business associate analysis applies and an agreement is required. For research use, which is a large share of the activity here since the products carry an explicit investigational disclaimer and are not cleared for diagnostic use, the governing instruments are different: institutional review board oversight and either an authorisation for research use or a waiver, not a services agreement. For pharmaceutical partners, manufacturers are generally not covered entities at all, so the operative instrument is the patient's own authorisation rather than any agreement the vendor signs.

Those three regimes do not overlap neatly and nothing published distinguishes them. A buyer should establish which one its own engagement sits under, and specifically whether studies contributed for clinical interpretation can be used for the research and biopharma side of the business.

A fourth question sits alongside them. The company's published academic collaborations are concentrated in Germany, Switzerland, Italy and France, so European data protection law reaches part of this estate whatever the position is in the United States, and nothing addresses it.

No attestation was located, and the site footer carries a privacy policy, terms of service and a cookie policy with no security page and no trust centre. That was checked against the footer directly rather than by search alone.

Company stage is a genuine part of the explanation here in a way it is not for larger vendors. This is a 2024 founded company on a single seed round, and an early stage business will not usually have completed an attestation cycle.

Stage does not dispose of the question, though, because of the operating model. This is not software installed inside a customer's environment. It is a hosted platform that receives whole body imaging studies uploaded from institutions, holds them for processing, and returns reports, which places the company in custody of identifiable patient imaging from the first deployment regardless of headcount or funding. The European academic institutions it works with will ordinarily require a data processing agreement and documented security measures before imaging moves, so a version of this material may already exist even though nothing is published.

Ask what security programme exists today, whether an attestation is underway and on what timeline, and what contractual security commitments the company is prepared to make in the interim. An early stage vendor that answers those three plainly is a better risk than one that publishes a badge and cannot explain its scope.

Retrieval caveat: the company's terms of service page is indexed by search engines but did not resolve when fetched directly, so material published only there was not reachable.

The company deserves credit for stating this plainly where many would obscure it. Its own product page carries the disclaimer that SelectPSMA is an investigational tool, has not been cleared or approved by the FDA or any other regulatory body, and is not intended for clinical diagnostic use. Separately, the company has publicly named achieving FDA clearance for its detection and selection products as a forward looking objective rather than an accomplished fact.

A buyer should read the entire suite as pre clearance and confine use to research and investigational contexts until that changes. The grade reflects candid disclosure of an unfavorable status, not regulatory achievement.

No governance framework or bias evaluation was located. The Medical Advisory Board is a clinical oversight structure rather than an AI governance programme, and no disclosure addresses performance variation across scanner types, imaging protocols or patient populations, all of which are known sources of drift in imaging AI.

One credit first, because it is unusual and this index looks for it specifically. The company publishes an edit burden figure, stating that over half of cases require no manual adjustment. Read plainly that also says up to half do require adjustment, and most vendors decline to publish the number at all. Publishing it is candour and it belongs on the record.

Against that, the headline performance claim is undefined. A figure of 92 percent is given as accuracy in identifying lesions, attributed to internal studies, with no split between sensitivity and specificity, no denominator, no definition of what counts as a correct identification and no external validation. Accuracy as a single number is not interpretable for a detection task where lesion counts vary enormously between patients.

The sharpest unanswered question is specific to this modality. The company states the product supports any tracer and is tracer agnostic. Different agents have materially different biodistribution and different patterns of physiological uptake in normal tissue, which is precisely what a lesion detector has to separate signal from. Tracer agnostic performance is therefore a strong claim that would need validation per agent, and none is published. Ask for performance broken down by tracer, by scanner manufacturer and by disease burden, since a model tuned on high burden metastatic studies may behave differently on early recurrence.

Products are described functionally and the imaging modality and clinical context are specific, and no model architecture, training data provenance, evaluation methodology, performance figures or warranty, indemnity or remediation commitment was located, beyond references to preliminary multicentre results with no accompanying detail. One characteristic of the evidence presented deserves naming because it recurs and is easy to mistake for support.

The published scientific citations relate to the underlying clinical problem rather than to the models themselves, so a reader encountering a reference list finds work establishing that the condition is important, that quantification matters, or that the modality is informative, and nothing establishing that this product measures it correctly.

Citing literature about the disease is not citing evidence about the tool, and a page dense with references can look better substantiated than a page with none while offering a buyer exactly as much. The distinction is worth applying wherever a vendor's bibliography is longer than its results section.

The clinical context sharpens what is missing, since quantitative nuclear medicine readouts inform treatment response assessment, where a measurement that drifts between timepoints changes a management decision without anyone seeing an error. Ask for repeatability on repeat scans of the same patient, agreement against expert reading, and the multicentre results with their methodology.

For an imaging analysis product this axis has to be read as imaging interoperability rather than chart integration, since the practical surface is the archive and the reporting pathway rather than the electronic health record. Read that way there is something here, but very little detail.

What is stated: studies can reach the platform by individual upload, by bulk upload or through PACS integration, and outputs come back as a structured report including staging classification, with structured export to support biomarker, response modelling and radiomics work. That combination fits a nuclear medicine and research workflow reasonably well, and bulk ingestion is genuinely useful for retrospective cohort work.

What is missing is everything that would let a department judge the integration before buying. No archive or reporting platform is named, no version support is stated, no interoperability standard is named anywhere despite this being imaging where the relevant standards are long settled, and nothing describes whether results return to the archive as a structured object or only exist inside the vendor's own interface. Nothing addresses the chart at all, so a finding generated here reaches the oncologist through the report rather than through the record.

Ask which archives are supported, whether output is written back as a structured imaging object, and how a result reaches the treating clinician who is not the reader.

The hosting model is now clear from the company's own workflow description even though no formal disclosure exists. Studies arrive by individual upload, bulk upload or PACS integration, analysis runs on the company's platform, and the product is described as cloud based with a hosted application login. So the answer to the question a hospital privacy office asks first is settled: imaging leaves the institution on every case.

That matters more here than in most of this lane because of what the imaging is. Whole body PSMA PET/CT covers the patient from skull to thigh, and cross sectional imaging that includes cranial anatomy is among the most re identifiable material in medicine, since a face can be reconstructed from it. This is not a cropped region of interest being sent for analysis.

What is still missing is everything a contract would turn on: no cloud provider, no processing or storage region, no residency commitment, no tenancy model, and no statement of whether an on premise or in tenancy option exists for institutions that cannot send studies out. The company's published collaborations are heavily European, which makes region a live question rather than a formality.

One inconsistency to resolve while asking: the site states analysis takes 7 to 10 minutes per scan in its workflow description and under 60 seconds in its summary statistics. Those are different claims about the same product.

No published pricing and no described commercial model. The buyer set spans hospitals, research institutions, imaging OEMs, and pharmaceutical partners, and each of those would carry different commercial structures, none of which are public.

Deliberately narrow and explicit about it: prostate cancer via PSMA PET/CT, with the products spanning detection, patient selection for radioligand therapy, and response monitoring across that single pathway. The company states an intent to expand beyond prostate cancer into broader oncology and advanced imaging, which should be treated as roadmap rather than current capability.