Radiology AI triage and flagging

A quality result is a flag that names the study and the visual finding, or an empty field. The model pass on CT, MRI, and X-ray exists to reorder the radiologist worklist. It does not close the examination.

Write the flag so the next person can open the same study and look at the same pixels. That means a study identifier (accession number or PACS study instance UID), plus the finding in visual language: modality, laterality, series, approximate slice or region, and the appearance the model used. Attach a priority score operations can sort on. Do not attach an impression, a diagnosis code, or a draft report.

Vendors in this class (Aidoc, Viz.ai, and the worklist and viewer surfaces that sit next to Epic and Microsoft imaging stacks) differ in screens and notification paths. The operations contract does not. If the pixels do not support a critical finding you can cite, the flag stays empty. Empty is a completed run.

The radiologist still reads every study. A high score moves a case up. A missing flag does not mean the study is normal. It means this pass did not assert a critical visual finding you can point to.

Start from the study, not from the score. Pull the examination the same way a radiologist would: correct patient, correct accession, complete series set, and a viewer that can show the slices the model claims. If the study is incomplete, burned-in laterality is ambiguous, or the series the model named is missing, stop. Do not flag a volume you have not loaded.

Once the study is on screen, allow the computer vision pass to propose a critical. Accept a proposal only when you can cite it: this study ID, this series, this region, this appearance. Then write the flag in that order and assign a priority score for worklist ordering.

Example, not a performance story. A night CT head lands at 02:10. The pass proposes a high priority. The usable flag is: accession ACC-441982, noncontrast CT head, series 3, slices 18-22, hyperdense extra-axial collection along the right convexity. The worklist places that examination first. The radiologist opens those slices, confirms or refutes the collection, and dictates the report. A later two-view chest for cough produces no visual critical. The flag field stays empty and the study keeps arrival order. That empty result is correct.

Do not let the score exist without the cite. A "stat" or "critical" rank with no study ID and no region is not a triage result. It is a pager with no body. Downstream staff cannot jump to the finding, and the radiologist cannot tell whether the model looked at this examination or a different one.

After the flag is written, the radiologist still reads. The flag is a pointer into the pixels. The report is the interpretation.

Most examinations will not carry a critical visual finding this pass is allowed to assert. For those studies, the correct output is an empty flag and no elevated priority.

Empty is not a failure to detect disease. The model is not being asked to call every abnormality, grade every nodule, or certify a normal. It is being asked to surface a small set of time-sensitive visual findings that change worklist order. If the study is ordinary relative to that bar, write nothing.

Do not invent a critical to keep the field busy. Do not write "possible bleed," "rule out ICH," or "cannot exclude PE" when the pixels do not show a finding you can cite. Those phrases are not flags. They are hedges that fill the worklist with work that is not there.

If a later human read finds a critical the model missed, that is a false negative for operations to review. It is not a reason to backfill a flag after the fact so the log looks complete. The historical record should show what the pass actually asserted at the time the worklist was ordered.

Three failure modes show up as "the AI is live" while quality is already broken.

A priority with no finding. The worklist shows a high rank, a red badge, or a critical column, and the flag text is blank, "abnormal," or "see AI." Nobody can open the cited region because none was cited. Charge nurses and overnight radiologists then treat every badge as noise, or they treat every badge as gospel. Either way the ordering signal is gone. Stop this at the interface: no priority score unless the study ID and the visual finding are both present.

Treating the flag as the report. Someone copies the model text into the impression, into the EHR problem list, or into a handover note as if the examination had been read. The flag is a worklist signal. It is not a signed interpretation. If operations allow the flag to travel as a diagnosis, you have replaced the radiologist with a ranking. Keep the flag on the worklist and in the viewer overlay. Keep the report in the reporting system, after a human read.

Inventing a bleed. The dangerous empty-avoidance pattern is a hematoma, occlusion, or pneumothorax that is not in the pixels. It happens when a downstream step is told to "always return a critical," when a generic "intracranial hemorrhage" label is applied because the protocol was a trauma head CT, or when laterality and series are guessed so the field is not blank. That is a fabricated critical. Ordinary stays empty. If you cannot name the study and the visual finding, you do not have a flag.

Watch the same pattern on related diagnostic passes. A lab spike with no analyte and no specimen, or an early warning score with no cited vital or lab driver, is the same failure: a rank without a finding.

Radiology triage is one image-and-score pattern among several diagnostic flags operations already run. The same rule applies on adjacent work: cite the evidence or leave the field empty, and do not let the score replace the clinician.

pathology image analysis is the closest cousin on still images of tissue: a region and a visual finding, or silence, with the pathologist still reading the slide. clinical photography classification is the same discipline on photographs rather than cross-sectional series; a label with no lesion location is as unusable as a CT flag with no slice. lab result anomaly detection is the non-image version of "flag or empty": an analyte, a specimen, and a reason, or nothing. early warning score automation is the other priority-score trap: a rank that fires without a cited vital or lab will train the floor to ignore the number, the same way an uncitable radiology badge trains the worklist to ignore red.

Keep those contracts aligned. A hospital that allows an invented bleed on CT while insisting labs stay empty on ordinary results will not have a coherent quality bar. The radiologist still reads the study. The flag, when it exists, only says which examination to open first and where to look.