HOPPR, a Chicago-based company focused on medical imaging AI development, has released the HOPPR® EB 2D Mammo Narrative Model, a vision-language model (VLM) that generates structured narrative language from 2D mammography images. Trained on more than 200,000 mammogram studies from multiple U.S. sites, the model is designed as a software component for developers building AI-assisted breast imaging and radiology workflow applications.
HotSpot Take
HOPPR has launched a vision-language model that translates 2D mammography images into structured narrative language, expanding its growing portfolio of radiology-focused VLMs. The HOPPR® EB 2D Mammo Narrative Model is trained on more than 200,000 mammogram studies from multiple U.S. sites, covering varied breast density categories and implant-displaced imaging scenarios. Accessed through the company’s Forward Deployed Services team, the model is intended as a foundational component for developers building AI-assisted breast imaging applications, not as a clinical end-product. It follows HOPPR’s recently released chest radiography narrative model and reinforces the company’s focus on giving developers configurable, traceable building blocks for medical imaging AI.
A Developer-Focused Tool for a High-Stakes Clinical Setting
A radiologist reviews mammography images at a diagnostic workstation. AI-powered narrative generation tools aim to reduce documentation friction in breast imaging workflows.
Breast cancer remains the most common cancer in U.S. women, with the American Cancer Society estimating approximately 321,910 new invasive cases in 2026. When detected at an early, localized stage, the 5-year relative survival rate exceeds 99%, making screening accuracy a clinical and public health priority.
Radiologists interpret more than 40 million mammograms in the United States each year. A large-scale prospective study published in Nature Medicine, involving more than 463,000 women and 119 radiologists across Germany, found that AI-supported screening detected breast cancer at a rate 17.6% higher than standard double reading, without increasing recall rates. The research adds to a growing body of evidence that AI has a measurable role to play in mammography workflows.
HOPPR, founded in 2019, approaches the AI development challenge from the infrastructure layer. Rather than building clinical applications directly, the company develops tools and platform capabilities that allow other developers to build those applications more efficiently and with greater traceability.
Narrative Language as a Workflow Building Block
The HOPPR® EB 2D Mammo Narrative Model generates structured JSON output from standard 2D mammography images, providing a language layer that downstream radiology workflow applications can consume. The training dataset covers diverse screening presentations, including varied breast density categories and implant-displaced imaging scenarios, across multiple U.S. sites.
“This model gives developers a practical foundation for building breast imaging applications that can generate structured language from images while still being configured, validated, and governed for their specific environment.” — Khan Siddiqui, MD, Co-Founder and CEO, HOPPR
“Mammography is one of the clearest examples of where AI needs to fit into existing imaging workflows, not force teams to rebuild them,” said Khan Siddiqui, MD, co-founder and CEO of HOPPR. “This model gives developers a practical foundation for building breast imaging applications that can generate structured language from images while still being configured, validated, and governed for their specific environment.”
The model includes version control that allows development teams to lock specific model versions, supporting consistency as applications are built and updated. According to the company, training data records are maintained to support traceability and bias assessment across the model lifecycle.
Access Through Forward Deployed Services
The HOPPR® EB 2D Mammo Narrative Model is accessed through HOPPR® Forward Deployed Services (FDS), the same channel used for its chest radiography narrative model. The FDS team works with partners to evaluate the model and configure integrations for their specific use cases, workflows, and data environments.
That delivery model reflects HOPPR’s view that medical imaging AI components are not plug-and-play tools. Each deployment involves clinical context, workflow variation, and governance requirements that affect how a model should be configured and validated. The Forward Deployed Services structure keeps that customization work as part of the implementation process rather than leaving it to the developer alone.
Portfolio Expansion and the VLM Strategy
The 2D mammography model follows HOPPR’s recently released chest radiography narrative model, which applied the same VLM approach to chest X-ray images. Together, the two models represent a pattern: HOPPR is building out a library of modality-specific narrative capabilities on top of its HOPPR® AI Foundry development platform.
The company also recently announced that NVIDIA’s open models, NV-Reason and NV-Generate, are available on the HOPPR® AI Foundry, expanding developer access to reasoning and generative AI capabilities for medical imaging.
Vision-language models occupy a specific position in the radiology AI landscape. Where earlier generations of imaging AI produced classification labels or bounding boxes, VLMs generate natural language output that can feed into reporting workflows, clinical decision support tools, and downstream documentation systems. The ability to produce structured narrative from images, rather than requiring a radiologist or application to translate AI output into prose, addresses a practical friction point in workflow integration.
What This Means for Breast Imaging AI Development
For developers working on breast imaging applications, the model’s value proposition is straightforward: a pre-trained, traceable VLM foundation that reduces the time and data requirements needed to build narrative generation capabilities from scratch. The 200,000-study training dataset, drawn from multiple U.S. sites, provides coverage that most individual organizations would not be able to replicate internally.
The broader significance lies in HOPPR’s modular approach. By releasing targeted VLMs for specific imaging modalities, the company is enabling developers to assemble AI-powered imaging applications from validated components rather than building end-to-end from raw data. That approach has implications for both development speed and the governance trail required for regulatory preparation.
For health systems and imaging vendors evaluating AI development strategies, the release signals continued maturation of the medical imaging AI development market, where the emphasis is shifting from demonstrating that AI can work in imaging to building the infrastructure that makes it work reliably at scale.
— This original article was created with AI support.
