HOPPR, the Chicago-based company building AI infrastructure for medical imaging, has released a vision-language model specifically tailored to mammography. The HOPPR EB 2D Mammo Narrative Model converts two-dimensional mammography images into narrative descriptions and structured JSON outputs, ready to plug into reporting systems, RIS, PACS, and triage workflows.
What the model actually does
Unlike the microcalcification and lesion detectors already on the market, the new component is not pitched as a viewer plugin. It is offered as a building block for developers who are constructing breast imaging applications. The input is a standard 2D mammogram, the kind used in population screening programs, and the output is a short, report-style description paired with a JSON payload that captures findings, categories, and references for any downstream system.
HOPPR cofounder and CEO Khan Siddiqui, MD, summarized the pitch: mammography is one of the clearest cases where AI must fit into existing workflows rather than force teams to rebuild them. The model gives developers a practical foundation to add structured language generation to their products without training a network from scratch.
Training set and data diversity
According to HOPPR, the model was trained on more than 200,000 mammography studies collected from multiple U.S. sites. The sample includes images across different breast density categories and challenging scenarios such as implant-displaced views. Sample diversity is a sensitive topic in breast imaging: recent studies have shown that algorithms trained only on specific populations can lose sensitivity in dense breasts, exactly the subgroup with the highest hidden risk.
The company also emphasizes that complete training-data records are maintained. Those records support auditability, bias assessment, and lifecycle traceability. In a field where the FDA has already engaged on density disclosure and screening equity, that kind of governance is becoming a commercial requirement, not a nice-to-have.
Version control and clinical governance
The HOPPR EB 2D Mammo Narrative Model ships with mechanisms that let clinical and IT teams lock a specific version of the model during deployment and updates. In other words, an institution can validate one version for clinical use and, even when HOPPR releases a new iteration, keep the approved version in production until a fresh validation cycle is complete. That discipline maps onto the FDA recommendation for Predetermined Change Control Plans in AI-enabled devices.
Version pinning also helps retrospective research: investigators can reproduce results because they know exactly which model version was used in any given analysis. For academic groups, where mammography datasets are increasingly used for theses and trials, that traceability is a meaningful differentiator.
The foundation model angle
HOPPR has been building a portfolio of foundation models for medical imaging through its AI Foundry platform. Before the mammography release, the company introduced a chest radiography narrative model. It also partnered with NVIDIA to make the NV-Reason and NV-Generate open models available inside Foundry, expanding the reasoning and text-generation toolset for imaging.
The positioning mirrors a wider trend across the sector. As we explored in our analysis of the five critical questions every radiology director should ask before adopting AI, platform strategies win because they let different services — breast, chest, abdomen, neuro — be served by specialized models without fragmenting infrastructure. For teams already exploring breast radiomics, it is worth revisiting the ISMRM 2026 work on PD-L1 prediction in breast cancer by MRI, which previews the kind of structured output that downstream apps can consume.
What this means for clinical practice
For breast imaging services, the practical impact will not arrive as a button on the radiologist workstation overnight. It surfaces in deeper layers: structured reporting applications, automatic triage of negative exams, summary generation for referrals, and quality audit databases. Clinical IT teams should prepare for vendor conversations in the coming months, as software providers begin to offer integrations built on top of this kind of model.
For research groups, the JSON output is especially valuable. It can feed supervised learning pipelines, quality dashboards, and retrospective studies with automated finding extraction, removing the manual annotation bottleneck. In regions facing a shortage of dedicated breast radiologists, that automation can accelerate population audits and reduce time-to-insight for screening programs.
What is next
Access to the new model is provided through HOPPR Forward Deployed Services, a team that evaluates and configures integrations based on each partner workflow and data requirements. The format signals that, for now, the product is delivered as a tailored engagement rather than a marketplace package. For interested teams outside the U.S., the route runs through direct contact and a technical proof of concept before adoption.
Founded in 2019, HOPPR develops AI infrastructure and tooling for medical imaging. The mammography model widens a portfolio that already covers chest radiography narration and partnership-driven reasoning models. Decision-makers evaluating the offering should expect the typical due diligence around model performance on local populations, data governance, and integration cost — but the building blocks are now in place to make those conversations concrete.
Source: DOTmed News
