contextflow

contextflow's own models are the engine of its core product, and its suite now includes other companies' AI. ADVANCE Chest CT runs the company's own convolutional networks for nodule detection and tracking, lung tissue quantification, incidental coronary calcium and a 3D image search that retrieves similar cases from a curated knowledge base; none of that exists without the models.

The contextflow AI Solution Suite also sells RevealAI-Lung, malignancy scoring from RevealDx, and CINA-iPE, incidental pulmonary embolism detection from Avicenna, and a lung cancer screening workflow built with Mint Medical. Remove contextflow's models and the company would still sell partner algorithms and a screening workflow, so the models drive the core product without being the whole of what is sold.

The radiologist decides, the evidence behind each finding is shown, and the thresholds are not published. The FAQ states the technology supports radiologists during diagnosis and does not perform automated diagnosis. Findings appear inside the radiologist's own viewer, and the similar case search shows comparable cases from the knowledge base with links to differential diagnosis literature, so the reader can inspect why the system suggested a pattern and disagree with it; users describe it as a double checker that leaves the final decision to them.

Nodules are classified and measured against named guidelines. What is missing is calibration: no sensitivity, specificity or operating point is published for detection or pattern classification, and nothing says how the confidence of a similar case match is shown, which matters when a benign looking comparison could reassure a reader about a malignant nodule. Structure described, thresholds missing.

The method, the reference data and the versions are described; performance is not. The FAQ says contextflow develops and trains its own deep convolutional networks that build embeddings in which similar patterns sit close together, using a mix of unsupervised, semi-supervised and supervised learning with radiologist annotation. The knowledge base used for similar case retrieval is described (more than 8,000 anonymized lung CTs from partner hospitals over about three years).

Product versions are visible: the company publishes DICOM conformance statements for ADVANCE Chest CT from version 2.0 to 7.0. There is no description of the data the detection and classification models were trained on, no performance figure per capability and no model card, and nothing says how customers learn what changed between versions.

A short chain for contextflow's own product, with its partners named. The company develops and trains its own networks, and ADVANCE Chest CT runs on a virtual machine inside the hospital's network, so patient images are not transferred to a vendor cloud for processing. The knowledge base behind similar case search comes from named partner hospitals' anonymized data.

The suite also carries partner products that are named with their developers: RevealAI-Lung from RevealDx and CINA-iPE from Avicenna, with a screening workflow built with Mint Medical. The site names no subprocessors, and it does not say whether any part of the knowledge base is queried remotely or how the 4DMedical acquisition changes the chain.

A peer reviewed workflow study backs one claim; detection performance is not published. contextflow cites a clinical study with the Medical University of Vienna, published in European Radiology in July 2022, reporting that average report reading time fell by 31 percent with ADVANCE Chest CT, and it names users at Erasmus MC, Freiburg University Medical Center, the Medical University of Innsbruck, Leiden University Medical Center, the Medical University of Mainz and the NHS South West London imaging network.

The partner malignancy score it resells (RevealAI-Lung) rests on a 2022 study in the Journal of the American College of Radiology, which is RevealDx's evidence rather than contextflow's. No sensitivity, specificity or reader comparison is published for nodule detection, tissue quantification or pattern classification, so the claim to improve reporting quality has no measurement behind it. Published research short of independent validation, covering time rather than accuracy.

The data flows are described specifically; retention and training are only partly pinned down. The FAQ says patient cases are processed locally on a virtual machine in the hospital's network, that the algorithm uses image information only, with patient text overlaid locally, and that the knowledge base used for similar case search contains more than 8,000 lung CT scans with reports from about three years of clinical routine at partner hospitals, fully anonymized so they cannot be traced to individual patients; partner hospitals are not paid for the data.

On training, the company says it does not plan to retrain on a customer's local data at the moment, which is a current intention rather than a commitment. Not stated: how long the local server keeps images, and how knowledge base cases were consented. Safety engineering, meaning how a missed or misleading finding is handled as a safety event, is not published. Categorical commitments on local processing and anonymization, without the retention schedule.

The product's privacy design is described in its FAQ; the company's role and agreement are not. contextflow is a European vendor working under GDPR. Its FAQ says the software runs on a virtual machine inside the hospital's network to comply with GDPR and data security rules, that the local server receives and stores DICOM images forwarded from PACS, that contextflow staff reach that server over a secured VPN and SSH for maintenance and support, and that the reference knowledge base holds only fully anonymized cases from partner hospitals.

The website privacy policy covers site visitors only. No data processing agreement is published and the company does not state its role (processor under GDPR, or business associate for a US customer), even though its support staff can reach identifiable images. A substantive description of how the product handles data, with no role or agreement behind it.

Controls are described with nothing independent behind them. The FAQ describes a local server behind the hospital's firewall, HTTPS that can be configured (with self-signed certificates as a fallback), user and password access, and vendor support over a secured VPN and SSH. The company's ISO 13485:2016 certification covers its medical device quality system, not information security.

No information security certification, attestation, penetration testing statement or trust portal is published. Controls asserted with no outside check. The stakes are higher because vendor staff can reach a server that stores patient DICOM images.

The position is stated plainly for both markets, without the certificate details. ADVANCE Chest CT is CE marked, and the company holds ISO 13485:2016 certification for its quality management system through TÜV SÜD. For the US, the product page states that the nodule detection feature is under development as a standalone product and not yet cleared for clinical use, so no US clearance is claimed.

The device class and certified intended use under the EU regulation are not given on the product or FAQ pages, and no certificate or registration number is shown. The June 2026 acquisition by 4DMedical, which holds US clearances for its own products, may open a US route; nothing has been filed or announced.

Governance is described as practice, with the knowledge base characterized and no results by group. The FAQ answers the bias question directly: the knowledge base must cover the relevant patterns for a population, data from hospitals is curated rather than taken as reported, reference content is updated as disease descriptions change, and the algorithms are continually validated against overfitting.

It describes the knowledge base as more than 8,000 lung CT scans from about three years of routine care at partner hospitals, selected on technical quality only to avoid a biased sample. No performance by scanner, reconstruction kernel, age, sex or other group is published, and the knowledge base's countries, centers and patient mix are not listed.

Both matter here: interstitial pattern classification is sensitive to reconstruction settings, and a similar case search can only show presentations its knowledge base contains. Governance stated as practice with no evaluation behind it.

Mechanisms let a reader check the output, and the published terms stand behind the software rather than its findings. Each finding comes with comparable cases from the knowledge base and links to differential diagnosis literature, so a radiologist can see what the system matched and reject it, and the software does not diagnose on its own. contextflow's published general terms (AGBs) provide an acceptance test, the statutory warranty period from acceptance or delivery, and state that no warranty claims arise from brochures or marketing statements; that warrants the software against its specifications, not the accuracy of a finding. No detection performance is published to hold the product to, and no indemnity or remediation commitment attaches to a missed nodule. Mechanisms to challenge the output, nothing behind it.

Integration into the imaging workflow is specified, with named partners. contextflow publishes DICOM conformance statements for each version of ADVANCE Chest CT (2.0 to 7.0). Its FAQ describes the path: a local server with a DICOM receiver accepts studies forwarded from PACS, and a PACS plugin opens the matching case in the radiologist's native viewer by instance UID.

The lung cancer screening workflow is built with Mint Medical, a structured reporting vendor, and users at named European academic centers put the quantified results into the radiology report. No record system integration is claimed. A published standard with named deployments behind it.

On premise by design, with the data path described and one piece left open. The FAQ says the software could run as a web application but is hosted within the hospital's network to comply with GDPR and data security rules: a local virtual machine receives DICOM from PACS, processes cases locally, and serves results over HTTPS behind user and password access, with contextflow reaching the server over VPN and SSH for support. Patient images therefore rest inside the hospital.

What is open is the similar case knowledge base: the FAQ says the company links to a reference database that is kept current, without saying whether that database sits on the local server or is queried remotely. Option and residency stated, with part of the processing path unclear.

The licensing terms are public; the price is not. contextflow publishes its general terms of business (AGBs, in German), which set out a software license from contextflow as licensor, electronic delivery, a one week acceptance test, and the statutory warranty period running from acceptance or delivery.

It sells ADVANCE Chest CT alongside resold partner products (RevealAI-Lung, CINA-iPE) to radiology departments, through PACS partners and to pharmaceutical companies, and the route in is a demo. The fee is not given, nor whether it is charged per study, per site or per year, nor whether modules can be bought separately. The company was acquired by 4DMedical in June 2026, and pricing may move into the parent's structure.

Chest CT only, with coverage named and validation behind part of it. ADVANCE Chest CT covers suspected lung cancer (nodules of 4 to 30 mm, classified and measured against British Thoracic Society, Fleischner, Lung-RADS, RECIST and EU screening guidance, and tracked over time), interstitial lung disease and COPD (quantification of seven findings and 19 image patterns), and incidental coronary calcium; partner products add malignancy scoring and incidental pulmonary embolism.

Users are named at academic centers in Austria, Germany, the Netherlands and the UK, and the company is building a workflow for Germany's national lung cancer screening program with Mint Medical. Published validation covers reading time in one Vienna study. Performance by setting goes unreported, as does any US use, and the nodule detection feature is described as not yet cleared in the US.