When radiologists hear "AI triage," the mental model that often comes to mind is an automated second reader: software that looks at scans and renders findings. That is not what worklist triage does. Understanding the difference matters for departments evaluating whether a triage tool fits their workflow.
What the System Is Actually Doing
AI triage for chest CT operates on a narrower problem than diagnostic AI. The question it answers is not "what does this scan show" but "does this scan likely contain a finding that warrants being read before other cases in the queue?" That is a classification task, not a diagnostic one. The output is a priority signal, not a finding description.
For ImageAssist specifically, the process begins when a completed chest CT study transmits via DICOM to the triage system, which happens in parallel with the study reaching the PACS. The system processes the volumetric dataset, evaluates it against trained models for the finding categories it handles (intracranial hemorrhage, pulmonary embolism, pulmonary nodule burden in the Lung-RADS 3 and above range, large pleural effusion), generates an urgency score, and communicates that score to the RIS via API. If the score crosses the threshold for priority placement, the study moves to the front of the worklist before the radiologist has opened a single case in that session.
End to end, this runs in under 60 seconds from DICOM receipt under normal conditions in our pilot deployments. The radiologist's reading environment looks exactly the same; the only change is which study is at position 1.
The Four Finding Categories
Limiting scope to specific finding categories is a deliberate design choice, not a limitation to be apologized for. A system trying to prioritize every possible finding of interest would require a much higher false-positive rate to achieve useful sensitivity, and would push studies to position 1 so frequently that the prioritization signal loses meaning.
The four categories ImageAssist covers represent findings where time-to-read has direct clinical consequence and where the imaging characteristics are distinctive enough to train a reliable prioritization model:
Intracranial hemorrhage on chest CT: While standard chest CT protocol does not typically include the brain, certain trauma and expanded chest protocols do. When hemorrhage is present and the protocol covers relevant anatomy, fast radiologist review directly affects clinical response time.
Pulmonary embolism: CT pulmonary angiography is the definitive diagnostic study for suspected PE. Central and saddle emboli especially carry time-sensitive clinical implications. These studies benefit from priority placement regardless of when they arrive on the overnight or weekend queue.
Pulmonary nodules at Lung-RADS 3 and above: This category is different in character from the acute findings above. A Lung-RADS 3 nodule is not an emergency; it requires timely follow-up management and clear communication. The triage value here is preventing these cases from sitting in a long queue where they may not receive the focused attention the follow-up workflow requires. We discuss this further in the dedicated pulmonary nodule prioritization article.
Large pleural effusion: Significant effusions, particularly those developing rapidly, can affect respiratory status and clinical management decisions. The triage model flags effusions meeting size thresholds for priority consideration.
What the Radiologist Sees
The worklist interface the radiologist uses is unchanged. If the department uses Sectra, Intelerad, Philips IntelliSpace, or any other major PACS, the radiologist continues working within that system. ImageAssist does not add a separate reading interface, a secondary viewer, or a pop-up overlay in the reading workflow. The only visible output is a priority indicator on the worklist entry for flagged studies, visible as a badge or status marker depending on how the RIS renders it.
When the radiologist opens a prioritized study, there is no AI-generated pre-read or auto-populated finding text. The radiologist reads the scan as they would any other, applies their clinical judgment, and dictates or otherwise generates the report in the normal fashion. The AI contributed to getting the right study in front of the radiologist first. Everything after that is clinical radiology.
Where the Radiologist Stays in Control
Radiologists sometimes ask whether the prioritization is binding: what if the AI flags a study as urgent and the radiologist disagrees once they open it? The answer is that every prioritization can be overridden. A radiologist who opens a flagged study, finds it unremarkable, and wants to continue working the queue has that option. The AI is providing a triage suggestion, not mandating a read order.
This matters clinically for a subtler reason. Triage models are trained on imaging characteristics and will produce false positives: studies flagged as potentially urgent that turn out to be negative on read. The false positive rate is a performance metric that departments deploying ImageAssist can review and adjust thresholds for. What false positives do not do is prevent the radiologist from reading other cases; they add a brief additional read at priority position that the radiologist handles normally. The cost of a false positive is a few additional minutes. The cost of a false negative, a truly urgent study that does not surface quickly, is the reason the system exists.
We are clear that this is a decision-support tool operating in a radiologist-in-the-loop model. That is not a marketing hedge; it reflects what the system does and what it does not do. The clinical evidence page describes how we think about validation in this context.
Integration in Practice
Deploying chest CT triage in a department requires two technical touchpoints: a DICOM listener that receives study transmissions, and an API or HL7 connection to the RIS for worklist update. Most major RIS vendors support worklist priority updates via HL7 ORM messages or proprietary APIs. The DICOM layer is protocol-standard and does not require modification to the scanner or PACS configuration.
A typical deployment in our experience involves an onboarding session to confirm the integration endpoints, a test phase using a de-identified case set to verify the DICOM transmission and worklist update are functioning correctly, and a go-live period where the system runs alongside normal workflow. The department's PACS administrator and IT team are involved primarily in the initial endpoint configuration; ongoing operation does not require dedicated IT overhead.
If you want to walk through how this would work with your specific RIS and PACS setup, the clinical team can discuss the integration specifics directly.
