Nucs AI

The model is the product. Nucs AI describes an end to end deep learning model that reads PSMA PET/CT. DeepPSMA detects lesions, assigns TNM stage and quantifies whole body tumor burden; SelectPSMA predicts response to PSMA targeted radioligand therapy; TrackPSMA compares scans over time to evaluate treatment response. The platform around them handles upload, physician review and report generation.

The same outputs are sold to pharmaceutical companies, contract research organizations and researchers as trial and biomarker data. Remove the models and nothing is left to sell.

The oversight structure is described. Every case goes through physician review: the reader checks and validates DeepPSMA's lesions, staging and tumor burden in the interface and can edit them before any structured report is generated. Nucs AI says more than half of cases need no manual edits. The products are labeled investigational and not for clinical diagnostic use.

One part is missing: what happens to what the reader does not notice. No figure shows how often physicians change the automated stage, or how often a missed lesion is found later. The homepage says AI driven quantification "ensures reproducible, consistent reporting across institutions". That is a quality claim which needs agreement data between sites and between readers, and none is published.

The architecture is described in general terms: an end to end deep learning model processes each PSMA PET/CT scan in 7 to 10 minutes (the homepage elsewhere says under a minute). The outputs are named precisely: lesion detection, TNM staging, tumor volume and SUV values, response prediction and response evaluation, with a CSV export for research.

Nothing is identified beyond that: no architecture, training data size or source, version, or update practice. Nucs AI names Segmed, an imaging data aggregator, as a data partner for biomarker development, which may bear on where training data comes from.

Named parties are published. The privacy policy lists Google Cloud and Auth0 as third party service providers for the platform, and the models are Nucs AI's own deep learning, with no outside model provider named in the path. Nucs AI also names Segmed, an imaging data aggregator, as a partner for biomarker development and real world evidence.

The list is short and its terms are not published, and because the products are not cleared the model can change without regulatory review. Nothing says whether customer studies flow to Segmed or pharmaceutical partners, or what consent covered the training data.

The performance figures published are the vendor's own, without method or denominator. The DeepPSMA page cites internal studies for 92 percent accuracy in detecting PSMA positive lesions, 87 percent in TNM staging and 99 percent in tumor quantification. For response prediction it cites a 94 percent negative predictive value for identifying non responders, and 4.56 times the risk of progression and 2.1 times the risk of death for predicted non responders. It lists an ASCO GU 2025 poster, an RSNA 2024 oral presentation and a 2024 Journal of Nuclear Medicine paper (Gafita et al.) on outcome prediction after Lu-PSMA-617.

None of the product figures gives case counts, reference standard or study design on the page, and accuracy as a single number does not separate missed lesions from false ones.

The privacy policy (effective July 2024) covers the website and the platform, and its assurances are general. Health information is "retained as per legal requirements or until no longer necessary". Data is stored in encrypted form and access is limited to staff who need it, with an access log. Breaches are notified within 36 hours. The same policy allows disclosure of "aggregated, de-personalized information" to other parties "without restriction, such as for marketing, advertising, or other uses".

Not stated: how long uploaded PET/CT studies are kept, whether customer scans train the models, and how de-identification is done before data is aggregated or shared with partners such as Segmed. Whole body PET/CT includes the head and is highly identifiable.

The privacy policy reaches the platform and addresses HIPAA, but states the wrong role for a vendor selling to hospitals. It says Nucs AI, Inc. is a "covered entity" under HIPAA and attaches a HIPAA Notice of Privacy Practices, the document a care provider gives its own patients. A company analyzing scans for hospitals would ordinarily be their business associate, and no business associate agreement is mentioned or offered. The policy also addresses EU, Brazilian and Canadian users.

The products are investigational and not cleared, so much use is research, where IRB approval and patient authorization matter more than a services agreement. Pharmaceutical partners are usually not covered entities at all.

Controls are asserted with nothing independent behind them. The privacy policy says data is kept on secure servers behind firewalls, stored encrypted, restricted to staff who need it with an access log, and that breaches are notified within 36 hours. No security certification, audit, penetration test or trust center is named, and the footer has no security page.

The platform holds whole body PET/CT studies uploaded from other institutions, on Google Cloud with Auth0 for sign in (both named in the privacy policy).

The status is stated plainly and it is pre clearance. The DeepPSMA page says the product is an investigational tool, not cleared or approved by the FDA or any other regulator and not intended for clinical diagnostic use. The FAQ adds that CE marking and FDA clearance are both in progress. SelectPSMA and TrackPSMA carry the same investigational position, and access is by request for investigational use.

The homepage still describes clinical uses, such as helping physicians "stage, quantify, and decide with confidence" and a "Clinical Diagnostics" use case, which the disclaimer rules out for now. Until a clearance number exists, the whole suite remains a research and investigational offering.

Robustness is asserted with no groups or counts. The DeepPSMA FAQ says the product "supports any PSMA tracer, it is PSMA agnostic", and the homepage promises consistent reporting across institutions. PSMA tracers differ in how they distribute through normal tissue, which is what a lesion detector has to separate from disease, so that claim needs results by tracer. None are published.

No results are given by tracer, scanner maker, site, disease burden or patient group, and no governance framework or bias evaluation is described. The edit rate is published: more than half of cases need no manual change, so up to half do.

A physician can challenge each output before it is used. Findings are reviewed and validated in the interface, with edits allowed before a structured report is generated.

Nothing stands behind the output. The terms of service disclaim any warranty on the completeness, accuracy or reliability of the platform and cap damages at the fees paid, and the customer indemnifies Nucs AI. The accuracy figures come from internal studies with no case counts or method, so there is no error rate a buyer can hold the company to.

Integration is claimed without anything named. Studies reach the platform by individual upload, bulk upload or "PACS integration", and DeepPSMA claims "seamless radiology integration". Results come back as a structured, shareable report with staging, and research users can export structured data as CSV.

No PACS, reporting system or record system is named, no standard is described, and nothing says whether results return to the archive or stay inside the Nucs AI interface.

One hosted option is described. Studies are uploaded to Nucs AI's cloud platform (sign in at app.nucs.ai) and analyzed there. The privacy policy names Google Cloud as a service provider and Auth0 for authentication. It says transfers to a third country follow the user's documented instructions and data is stored encrypted.

No region or residency commitment is published, no on premises or customer tenancy option is described, and tenant isolation is not stated. Whole body PET/CT leaves the institution on every case.

No price is published, but who it is sold to and how access starts are visible. The site addresses clinicians, biopharma teams, researchers and strategic partners. Access to DeepPSMA for investigational use is by request through a contact form, and the terms of service refer to a fee "if any" for platform access. Three products are sold separately (DeepPSMA, SelectPSMA, TrackPSMA).

Coverage is stated clearly and narrowly: prostate cancer on PSMA PET/CT. It spans detection and staging, prediction of response to PSMA targeted radioligand therapy, and response evaluation, for clinical research, trial imaging and radiomics. The products are investigational and not for clinical diagnostic use.

Nothing published yet validates the stated coverage with counts: the accuracy figures come from internal studies with no populations described, and the any tracer claim has no results by tracer.