Radiology Triage screens every incoming study the moment it lands in PACS, flags suspected critical findings, and reorders the worklist so the bleed is read before the routine follow-up. Radiologists make every call; the queue just stops being first-come, first-served.
Most worklists are ordered by arrival. A head CT with an intracranial bleed sits behind a routine knee MRI because the knee got there first — and nothing in the queue knows the difference. The radiologist finds out when they open it.
The bleed is fourth in line. It waits for three studies that could have waited for it.
Same four studies, same radiologist, same shift. Only the order changed.
Three failure modes sit underneath this, and they are not the same problem. A finding can be missed on the image. It can be misjudged once seen. Or it can be seen correctly and never reach the physician who needed it. The third is the one software is best placed to fix, and the one most tools ignore.
Every study is screened on arrival for suspected critical findings — hemorrhage, pneumothorax, PE, fracture. A positive flag moves the case to the top of the worklist within seconds, not at the next batch read.
Cases are matched to the right reader: subspecialty, credentialing, site and shift. The neuro CT reaches the neuroradiologist on call, and the queue balances itself across the group.
A reader who disagrees with a flag, or with peer review feedback, can request a second opinion from another radiologist. The disagreement gets resolved on record instead of quietly dropped.
Every flag, reorder and override is logged with the model version that made it. When a reviewer asks why a case jumped the queue, the answer is on record — explainable, exportable, retained.
Flags land directly on the worklist your radiologists already read from — no second screen, no separate login. Shown here: the triage queue with a critical flag at the top.
Five areas of work. The first three are the platform we are building now; the last two are where the roadmap goes once the workflow underneath them is proven in production.
The queue stops being chronological. Studies are ordered by clinical urgency, matched to a reader who is credentialed for them and actually on shift, and surfaced with the reason they moved.
The part most tools leave out. When a reader disagrees — with a flag, with a peer reviewer, or with themselves — there is a route to an independent read, and the outcome is recorded either way.
A correct read can still produce a wrong report. Measurements are carried from the imaging system into the report rather than dictated from a worksheet, and the format is the same across the department.
Peer review that produces something. Cases are sampled and reviewed inside the same workflow, disagreements route to second opinion instead of ending in a spreadsheet, and the findings come back as something a department can act on.
Once the workflow underneath is solid, the same pipeline can carry quantitative work: features extracted from the image that sit below the threshold of visual assessment, combined with the clinical record.
In most departments a radiologist who disagrees with a peer review has no formal route to challenge it. The disagreement is absorbed, not resolved. Nobody learns anything and nothing is written down.
Radiology Triage makes the second opinion a first-class object in the workflow. A reader raises the request from the study they are already looking at. It routes to a second radiologist with the right subspecialty. Both readings are kept, the resolution is recorded, and the case is available to the department’s quality programme afterwards — which is where peer review was supposed to lead in the first place.
| № | Property | Value | Remark |
|---|---|---|---|
| 01 | Platform | Multi-tenant SaaS | The product. Hosted and shared, so a new site is a login rather than a server — and every tenant gets each improvement as it ships. |
| 02 | Tenancy | Isolated | Your studies, your users and your audit log stay yours. Nothing is pooled across tenants without a written agreement. |
| 03 | Interfaces | DICOM · HL7 | Reads from the PACS and RIS you already run. Results return to the same worklist. |
| 04 | Data | HIPAA | A business associate agreement is signed before any data moves. De-identified for model work. |
| 05 | Models | Versioned | Each release is logged and reversible, and you can see which version made any given call. |
| 06 | Also available | On-premises | Where imaging data cannot leave the network, we scope and build a tailored on-site deployment. |
On-premises and tailored work is scoped as its own engagement and delivered by the team that built the platform, not handed to a third-party integrator.
The same platform, sized differently. A single radiologist and a twelve-hospital system need the same thing from a worklist — they just arrive at it through very different doors.
Emergency and inpatient imaging, where urgency is highest and the cost of a delayed read is measured in outcomes. On-premises where the perimeter requires it.
Independent practices balancing subspecialty coverage against volume, with peer review obligations and no dedicated informatics team to build this themselves.
Throughput-driven outpatient work, where the win is a queue that keeps routine studies moving while still catching the one that isn’t routine.
Multi-site, multi-licence, around the clock. Routing on credential, state licensure and shift is the whole job, and it is the part spreadsheets do worst.
A single reader with no department behind them. Second opinion is the feature that matters most here, because there is no corridor to walk down.
Learners and researchers focusing on radiology reporting quality.
Every flag carries the region that triggered it and the model version that made the call. A reader sees why before accepting or dismissing it, so the queue order is something you can argue with.
Your studies are not pooled with other tenants’ to train shared models. If we ever want to use your data to improve one, we ask first, in writing, and you can say no and keep the product.
Performance is measured on your own case mix during deployment and reported per site and per model version. We do not carry someone else’s benchmark over to your department.
“I have read studies for twenty-three years, and spent the last few publishing on where AI gets medical imaging wrong. That is exactly why I want it in the worklist — as a second pair of eyes that shows its work.”
Dr. Okur trained in radiology in 2002 and has read across military medical centres, a rehabilitation hospital specialising in neurological and musculoskeletal imaging, and teleradiology. She held research fellowships at Loyola University Medical Center in Chicago and has thirteen peer-reviewed papers in Skeletal Radiology, European Journal of Radiology and Pediatric Radiology, among others. Her 2025 work at CVPR and ETRA examines where generative AI fails in medical imaging, and how a radiologist’s eye movements change when an image is synthetic. She sets the clinical direction for Radiology Triage.
The questions that come up in every security review, answered before you have to ask them.
On the multi-tenant platform, in our hosted environment, isolated per tenant — your studies, your users and your audit log are yours and are not pooled with anyone else’s. Where your policy does not permit imaging data to leave your network at all, we scope a tailored on-premises deployment instead. That is a separate engagement, and we will tell you honestly which of the two fits you.
Not without asking you first, in writing. If we want to use your data to improve a model, we come to you with a specific request, and you can decline and keep using the product exactly as before. Consent is not buried in the terms of service.
No. We read from the PACS and RIS you already run over DICOM and HL7, and results return to the same worklist your radiologists already read from. No second screen, no separate login, no migration project.
A radiologist accepts or overrides every flag, and the override is a first-class action rather than a workaround — captured with the reason, the reader and the rule or model version involved. Overrides are how we find out what needs fixing, so the product is built to make them easy rather than discouraged.
Yes. Every prioritisation, routing decision, override and second opinion is written to an immutable log with actor, timestamp and version. Any change in queue order can be reconstructed and exported as CSV or PDF for a quality committee, an insurer or a regulator.
Radiology Triage is a Zinniamor Tech LLC product, based in Chicago. Clinical direction comes from a radiologist with twenty-three years of reading experience and published research on the failure modes of AI in medical imaging — not from a team that has only read about the problem.
A 30-minute walkthrough with a clinical specialist, on a de-identified copy of your case mix. We will show the flags, the misses, and the audit trail — not just the highlight reel.