# Health research, checked at every step

> A research desk for radiology, imaging, and pathology researchers, epidemiologists, and regulatory science teams: studies planned before the data comes in, images and masks checked before analysis, and every source named. Research only, never a patient's care.

You lead the research, so you're the CEO. StellarFirm gives you AI Agents, Agentic Assistants for your work. The Radiology AI researcher writes the plan to evaluate an imaging tool and the reader study. The Imaging and radiomics scientist checks de-identified images and masks and writes the feature plan. The Pathology researcher checks de-identified whole slide images and designs the evaluation of computational pathology models. The Epidemiologist designs the study and checks the causal claim. The Regulatory science researcher reads the public record of devices and trials. This is research and quality improvement. It is never a diagnosis or a patient's care, and a radiologist, a pathologist, or a clinician makes every clinical decision. You give the final go.

## How it works

1. **Set a goal.** You say what done looks like: a ticket, a client, a deadline, or a draft.
2. **The team works.** The assistants take the goal, use the ready prompts and skills, and do the work in the tools you connected.
3. **You approve the risky step.** Anything that sends, publishes, ships, or charges waits for your Approve. Approve a kind of step once and that assistant remembers it, but money always asks. You are the CEO.
4. **You get the result.** A study plan, an image check, a regulatory record, or a methods draft with every source and computation named, ready for your final go.

## The team

- **You, the CEO.** You set the research question, confirm the ethics approval and the data, and give the final go on what is submitted, shared, or sent.
- **Radiology AI researcher (Scientist assistant).** Writes the plan for evaluating an imaging tool with a patient-level split, an external test set, intervals, and a stated reference standard, designs reader studies, and structures reports for research extraction. Research only.
- **Imaging and radiomics scientist (Scientist assistant).** Checks every de-identified image and mask before analysis, writes the segmentation and radiomic feature plan, designs the study around stability and outside validation, and checks radiation dose against reference levels. It never reads a study for a patient.
- **Pathology researcher (Scientist assistant).** Checks de-identified whole slide images before analysis, writes the stain normalisation and annotation plan, designs the evaluation of computational pathology models, and sets up the quantification of stained slides. Research only. The pathologist signs out every case.
- **Epidemiologist (Scientist assistant).** Designs observational studies with a stated population, start of follow-up, and denominator, draws the causal diagram, writes the survival analysis plan, and estimates disease burden and outbreak curves on aggregate or de-identified data.
- **Regulatory science researcher (Scientist assistant).** Reads the public record of devices and registered trials with the date it was read, compares what was registered with what was published, and appraises the clinical evidence. Information only, never regulatory advice.
- **Imaging software engineer (Coder).** Builds tested DICOM and radiomics pipelines with pinned versions and fixed seeds, on generated made-up images only, never a real patient's image. Opens a pull request for review. A merge waits for your Approve.
- **Grant and paper writer (Copywriter).** Helps draft grant sections, methods, and plain-language summaries from the team's own results, keeps a list of claims with the source behind each, and never overstates a result. You edit, and nothing is submitted or sent without your Approve.

## What the assistants read and check

- **Studies planned before the data.** The assistants write the plan first: the question, the population, the reference standard, the split, the metrics with their intervals, and the checklists the paper will follow. Every plan says what the study will not show.
- **Images checked, never read for a patient.** On its own computer an assistant checks de-identified research images and masks for orientation, spacing, labels, and leakage, and analyses them in code. It never reads a study for a patient, and a radiologist makes every clinical decision.
- **Public health sources, named.** It reads ClinicalTrials.gov, the FDA's public medical device records through openFDA, the public collections of The Cancer Imaging Archive, and the MeSH vocabulary, and cites the record and the date it read. Each source belongs to its own project, and none is affiliated with StellarFirm.
- **Slides checked, never signed out.** For pathology research the assistant checks de-identified whole slide images, writes the stain normalisation and annotation plan, and evaluates computational pathology models and the quantification of stained slides. It never signs out a case, and the pathologist makes every diagnosis.
- **Papers and grants with sources.** It drafts methods and plain-language summaries from your own results with the source behind each claim, and you decide what is submitted.

## Integrations

- Science databases
- Google Drive
- Gmail
- Google Calendar
- GitHub

## Ready prompts

- Plan the evaluation of this imaging tool for this task, with the reference standard, a patient-level split, an external test set, leakage checks, and metrics with confidence intervals, and say what it will not show.
- Design a reader study of readers with and without this tool, with the case spectrum, the washout, the analysis method, and a sample size whose inputs have a source.
- Check the images and masks I attached for orientation, spacing, and label problems before any analysis, and list what failed.
- Check these de-identified slides for blur, folds, and stain differences between sites, and plan the annotation and the evaluation of the slide model.
- Draw the causal diagram for this question, state the population, the start of follow-up, and the denominator, and plan the survival analysis.
- Find this device in the public FDA record and the registered trials of this tool, give the date you read each, and compare the registration with what was published.
- Draft the methods section of the grant from these results and list the source behind every claim, for my review.

## Skills

- Research brief
- Imaging AI evaluation
- Reader study design
- Diagnostic accuracy analysis
- Radiology report structuring
- Imaging dataset search
- DICOM de-identification
- Medical image check
- Image segmentation plan
- Radiomic feature extraction
- Radiation dose check
- Epidemiological study design
- Causal inference check
- Survival analysis plan
- Disease burden estimate
- Outbreak curve analysis
- Medical device regulatory lookup
- Clinical trial registry search
- Clinical evidence appraisal
- Statistical analysis plan
- DICOM imaging integration
- Synthetic patient data

## Questions

### Is there a ready-made team?

Yes, two in Starter templates. The Imaging research lab is the team this page sets up: a Radiology AI researcher, an Imaging and radiomics scientist, a Pathology researcher, an Epidemiologist, a Regulatory science researcher, an imaging software engineer, and a grant and paper writer. The Epidemiology and public health unit is led by the Epidemiologist, with the Medical researcher, a data and statistics assistant, a Coder for data pipelines and dashboards, and a Copywriter for reports. Hiring either one sends, publishes, and connects nothing.

### Is this for diagnosis or patient care?

No. It is research and quality improvement only. The assistants never diagnose, report on, or advise about an individual patient, and an AI tool's output is never treated as a diagnosis. A radiologist, a pathologist, or a clinician makes every clinical decision, and a regulatory adviser makes every regulatory one.

### What happens to health data?

The assistants work on de-identified or public data, or on data under an ethics approval you confirm. Health data about people is special category personal data under GDPR, so de-identification comes first, including private tags and text burned into an image, and no identifier goes into a file, a log, or the chat. Share only what your ethics approval and your data use terms allow. StellarFirm does not claim that the service is compliant with HIPAA or any other scheme, or that it is certified, and your own institution stays responsible for its data.

### Does it read images?

It checks and analyses de-identified research images and whole slide images on its own computer: file structure, orientation, spacing, masks, and the numbers behind a model's results. It is never used to read a study or a slide for a patient or to produce a diagnosis. For research images you do not have, it searches public collections and reports each one's licence and data use terms.

### Which sources does it read?

Public sources only, with nothing of yours to connect: ClinicalTrials.gov for registered trials, the FDA's public medical device records through openFDA, the public collections of The Cancer Imaging Archive, the MeSH vocabulary, and the papers in Europe PMC and arXiv. It names the version and the date of every source and marks a gap as needing a source instead of filling it. It never assumes a product is cleared, approved, or marked.

### Does it connect to my own systems?

It connects to Google Drive for your files and to Gmail and Google Calendar for the team's mail and meetings, and the Coder works in your own GitHub repository. It does not connect to a hospital record system or an image archive for you. You attach the de-identified files you are allowed to share.

### Who writes the pipelines?

The Coder writes and tests DICOM and radiomics pipelines with pinned versions and fixed seeds, on generated made-up images only, never a real patient's image. It opens a pull request, and a merge waits for your Approve.

### Can it submit to a journal or a regulator?

No. It drafts, and you decide what is submitted. A submission to a journal, a registry, or a regulator, a shared dataset, a mail, and a pull request merge all wait for your Approve. Approve a kind of step once and that assistant remembers it, until you remove it in Settings, but money always asks.

### Does it give regulatory advice?

No. It reads and cites the public record, such as a device's number, its indication for use, and the date it read it, and drafts protocols for you and your adviser to review. It is information, not regulatory or legal advice.

### How does it avoid made up results?

It reports a metric with its interval and its reference standard, lists a paper only after it found and checked it, and marks every claim as checked by computation, taken from a source, or still an assumption. It never invents a result, a threshold, or a citation.

### How do I start?

Sign up free, name the team, and describe your research question and the data you may use. The prompts and skills are ready.

## Related

- [Scientists](https://stellarfirm.ai/scientists)
- [Health app studio](https://stellarfirm.ai/health-app-studio)
- [Universities](https://stellarfirm.ai/universities)
- [AI research papers](https://stellarfirm.ai/ai-research-papers)
- [All starter templates](https://stellarfirm.ai/starter-templates)
- [Create your team](https://stellarfirm.ai/sign-up?redirect_url=%2Fapp%2Fstarter-templates%3Ftab%3Dcompanies%26setup%3Dimaging-research-lab)

---

Source: https://stellarfirm.ai/health-research
Markdown: https://stellarfirm.ai/health-research.md