CxrReportGen Premium

What is CxrReportGen Premium?

CxrReportGen Premium is a specialized medical-imaging model from Microsoft, served through Microsoft Foundry. Unlike a general-purpose conversational model, it is designed around a narrow clinical workflow: generating structured draft findings from chest X-ray images.

The model is multimodal because it processes visual medical images together with text-based clinical context. A typical request includes a current frontal chest X-ray in PNG or JPEG format. The caller can also provide an indication, technique information, comparison details, a prior image, and a prior report. The output is generated text containing draft findings that can be reviewed and edited by a qualified healthcare professional.

Microsoft describes the model as a closed-weight, fully managed, serverless service. This means users access an operated model endpoint rather than downloading or hosting the model weights themselves. The served model identifier is CXRReportgen-Premium.

Where it fits in Microsoft’s healthcare AI catalog

CxrReportGen Premium belongs to Microsoft’s Health and Life Sciences premium model offering within Microsoft Foundry. Its specialization separates it from general-purpose language or vision models that may be able to discuss images but are not specifically presented as chest X-ray reporting systems.

The model is listed as being in limited preview. Registration and eligibility approval are required, so access should not be assumed to be generally available to every Foundry customer. Microsoft documents base-model deployment in East US, East US 2, West US 2, Central US, and West Central US. Fine-tuning availability can vary by region.

The supplied model information lists a release date of June 2026. Because the model is in preview, deployment regions, access requirements, pricing arrangements, and operational limits may change as Microsoft updates the service.

How the model works

The required input is a current frontal chest X-ray. The image can be supplied as a PNG or JPEG file. Optional fields allow the model to use information that would normally be relevant when a radiologist interprets a study:

• Indication: the clinical reason for the examination.
• Technique: information about how the study was acquired.
• Comparison: a description of earlier studies or findings to consider.
• Prior image: an earlier image for visual comparison.
• Prior report: the report associated with the earlier study.

Microsoft’s deployment documentation states that a prior image and prior report must be supplied together. The model returns generated findings text and usage token counts. The documented API default for max_tokens is 450, but Microsoft does not publish a separate absolute maximum output limit in the available deployment information. Therefore, the default should not be treated as a confirmed hard ceiling.

The output is text rather than a new image, audio recording, or video. Although the service is intended to support structured reporting, the supplied model data does not identify a separate guaranteed JSON mode or provider-defined structured-output schema. Integrators should verify the current endpoint documentation before depending on a particular response format.

Capabilities and limitations at a glance

Several important specifications are not published: the context window, a hard maximum output length, and a conventional knowledge-cutoff date. This is expected for a focused imaging service that relies primarily on supplied images and clinical fields rather than functioning as a general knowledge model.

Strengths for radiology workflows

The main strength of CxrReportGen Premium is task specialization. It is aimed at a concrete workflow rather than broad conversation. A healthcare organization evaluating automated assistance for chest X-ray reporting can send the image and relevant context to a managed endpoint without building a general vision-language system from scratch.

The ability to include prior studies is particularly relevant to longitudinal imaging workflows. When both a prior image and its associated report are available, they can provide comparison context for the generated findings. This may make the output more useful than an interpretation based only on the current image, although the quality and clinical relevance of the result still require review.

Microsoft also reports sub-one-second inference for typical single-frontal studies. This is a provider-reported performance claim rather than an independent benchmark, and real-world latency can vary with service conditions, request handling, deployment region, and surrounding application code. Nevertheless, the claim indicates that the model is intended for interactive or near-real-time assistive workflows rather than long-running batch processing alone.

Fine-tuning is another important option. Microsoft documents customization of premium healthcare models, including LoRA-based fine-tuning. LoRA, or Low-Rank Adaptation, is a method for adapting a model using additional learned parameters rather than retraining every original parameter. In practice, this could help an institution align outputs with local terminology or reporting conventions, but it does not remove the need for validation, governance, and clinical oversight.

Pricing and speed-versus-cost trade-offs

Standard inference is listed at $0.00218 per image. The supplied pricing information states that generated output is included in the per-image inference price. Fine-tuning is priced separately per image per epoch, so organizations considering customization must budget for both training data processing and standard inference.

At the listed rate, the pricing model is straightforward for workloads where one request corresponds to one primary image. Actual project cost can be higher when applications repeatedly resubmit studies, include additional processing outside the model, or use fine-tuning. The supplied information does not provide a monthly subscription price, volume discount, or minimum spend.

The model’s reported sub-one-second inference and per-image pricing make it potentially attractive for high-throughput first-pass assistance. However, speed and low unit cost should not be confused with clinical autonomy. A fast draft still needs review, and a low image price does not account for integration, security, data governance, validation, monitoring, or specialist review costs.

Reasoning, coding, and tool support

CxrReportGen Premium is not presented as a general reasoning or coding model. Its editorial reasoning score in the supplied data is 2 out of 10 and its coding score is 1 out of 10; these are database evaluations, not scores published by Microsoft. They reflect the model’s narrow purpose and should not be interpreted as a clinical accuracy benchmark.

The model does not have documented web search, external tool use, function calling, or action execution. It should therefore be treated as an image-and-context-to-text service. If an application needs literature retrieval, patient-record queries, workflow actions, scheduling, or broad clinical question answering, those functions would need to be implemented separately and governed independently.

The model record also lists no streaming support. Since the expected output is a relatively short set of findings, lack of streaming may be less important than it would be for a long conversational response, but it can still affect user-interface design.

Safety and clinical limitations

CxrReportGen Premium is not intended for autonomous diagnosis, unsupervised clinical decision-making, final reports without radiologist sign-off, or unvalidated emergency and time-critical workflows. Microsoft describes the generated findings as assistive and requires qualified human review.

A draft may omit a finding, mischaracterize an image, misunderstand the clinical context, or express a comparison inaccurately. Prior-study support also depends on the quality and correct pairing of the prior image and prior report. The model should not be treated as an authority simply because its output is fluent or formatted like a report.

Organizations should validate the system on representative local data before using it in production. Validation should consider different acquisition conditions, patient populations, image quality, prevalence of findings, reporting styles, and failure modes. Human-review procedures should make it clear which parts of the output were generated by the model and which were confirmed or changed by a clinician.

When to choose CxrReportGen Premium

CxrReportGen Premium is a reasonable candidate when the primary requirement is managed, specialized chest X-ray report drafting in Microsoft Foundry. It is especially relevant for:

• First-pass findings generation for frontal chest X-rays.
• Radiologist workflow assistance where every draft receives qualified review.
• Structured extraction or normalization of imaging findings.
• Research evaluation of healthcare AI reporting workflows.
• Institutions that want to investigate customization with local reporting data.
• Applications where a per-image price and reported low latency are more important than broad general-purpose reasoning.

Another type of model may be more appropriate when the application needs broad medical question answering, multiple imaging modalities, long-form reasoning, web retrieval, tool use, conversational interaction, or autonomous workflow actions. A general vision-language model may offer wider input flexibility, but it may lack the same task-specific positioning and healthcare workflow documentation. Conversely, a conventional radiology information system or clinician-authored workflow may remain preferable when the organization cannot yet validate automated findings or provide consistent specialist review.

Bottom line

CxrReportGen Premium is a narrowly focused Microsoft healthcare model for turning chest X-ray studies and related context into draft radiology findings. Its differentiators are specialization, managed Microsoft Foundry deployment, support for prior-study context, fine-tuning availability, a listed price of $0.00218 per image, and Microsoft’s reported sub-one-second inference for typical single-frontal studies.

Its boundaries are equally important: access is limited-preview, the model does not provide documented tools or web search, important limits such as context length are unpublished, and all generated findings require qualified human review. It is best evaluated as an assistive reporting component—not as an autonomous diagnostic system or a replacement for a radiologist.