Clinical decision support is one of the more overloaded terms in healthcare AI. It is used to describe tools that do very different things: order entry checks that flag drug interactions, EHR alerts that recommend a missing lab test, and imaging AI that surfaces findings in a radiology read. The same phrase covers all of them, which means it communicates almost nothing useful about what a specific tool actually does.
For radiologists evaluating tools that claim the CDS label, the vagueness creates a practical problem: it is hard to understand what you are being asked to adopt, what clinical responsibilities change, and what does not change, until someone is specific.
What CDS Is Not (the Relevant Version)
In the context of radiology AI, clinical decision support is often confused with autonomous AI reads. The confusion is understandable. Both involve algorithms operating on imaging data. Both produce some kind of output about what the scan contains. The distinction lies in whether the algorithm's output is advisory or terminal.
A diagnostic AI that generates a finding report, assigns a diagnosis, and routes that output directly into clinical documentation without radiologist review is not CDS. It is automated interpretation. This model is not how responsible radiology AI works today, and it is not what ImageAssist does. It is also not what any FDA-cleared or FDA-cleared-pending radiology AI tool in clinical use does for primary imaging findings.
CDS in radiology means the algorithm produces a recommendation, a flag, a risk score, or a prioritization signal that informs a clinician decision. The clinician, specifically the radiologist, reviews the underlying data and makes the clinical judgment. The AI contributes information to that judgment. It does not substitute for it.
The Specific Type of CDS That Worklist Triage Represents
There are several distinct CDS functions that radiology AI tools perform. Understanding which function a given tool performs matters for workflow integration and for understanding the radiologist's role.
Finding detection with radiologist confirmation: The AI identifies a candidate finding in the image and presents it to the radiologist for review. The radiologist confirms or dismisses the finding and incorporates it into the report. Examples include pulmonary nodule CAD and fracture detection overlays. The AI contributes to the "finding" phase of the read.
Quantification and measurement assistance: The AI measures tumor size, calculates volumetry, tracks change over time. The radiologist reviews the measurements and incorporates them into the report. The AI reduces measurement burden and improves consistency.
Workflow prioritization: The AI evaluates the study content and adjusts the sequence in which studies reach the radiologist. The AI does not generate findings and does not contribute to the report. Its entire function is upstream of the read: determining which study should be read first. This is what ImageAssist does.
These three functions have different clinical implications, different integration requirements, and different performance metrics. Grouping them under the same "clinical decision support" label is accurate in a broad sense but unhelpful when a radiologist is trying to understand what a specific tool will change about their daily work.
What Changes When You Use Workflow Prioritization CDS
For a radiologist using ImageAssist, the change to daily workflow is specifically in what study appears at the top of the worklist at the start of a session. The reading process is unchanged. The reporting process is unchanged. The radiologist's diagnostic responsibility is unchanged. The only thing that has changed is the sequence in which studies reach the radiologist's desktop.
The prioritization suggestion is not binding. A radiologist who opens a flagged study, finds nothing urgent, and prefers to continue reading in arrival order can do so. The triage signal is the same kind of advisory information as a stat flag from the ordering clinician: it represents a data point about likely urgency, not a clinical mandate.
This is a meaningful distinction from CDS tools that operate within the read itself. A CAD tool that highlights a candidate nodule in the image is present at the moment of interpretation and the radiologist must actively engage with it, either accepting, dismissing, or modifying the candidate finding. Worklist triage operates before the read begins and influences only the question of which study to open next. It has no presence inside the reading session.
The Radiologist-in-the-Loop Principle in Practice
The radiologist-in-the-loop framing is used frequently in AI marketing and occasionally dismissed as a compliance hedge. It deserves to be taken seriously as a description of how responsible radiology AI actually functions, because the alternative, AI that operates outside the radiologist's loop, has both regulatory and clinical implications that the industry is not ready to absorb.
For worklist triage specifically, the radiologist-in-the-loop principle means three things:
First, every study that receives a priority flag is read by a radiologist. The AI's priority signal does not replace the read; it requests that the read happen sooner. The radiologist reads the study, evaluates all findings, and generates the report. If the flagged study turns out to be normal, the radiologist documents that. If it contains the finding the model suspected, the radiologist identifies and characterizes it.
Second, the radiologist can and should override the prioritization when clinical context warrants it. A radiologist who receives additional clinical information, a verbal notification from the ED that a different case is more urgent, or personal knowledge of a patient situation can and should reprioritize manually. The triage prioritization is one input among several.
Third, the AI system's performance should be tracked and reviewed. Tracking priority accuracy, false positive and false negative rates, and time-to-read improvement for targeted finding categories provides the feedback loop that allows threshold tuning and performance monitoring. This is part of responsible CDS deployment, not optional. The evidence page on this site describes our approach to this tracking for pilot deployments.
Where the Regulatory Boundary Sits
It is worth being precise about regulatory framing. The FDA distinguishes between software that is intended to be used as a diagnostic aid and software that provides information to a clinician to support their decision-making. The line between these categories has been an area of active regulatory development, and the specifics matter for how a tool can be marketed and what claims can be made about its clinical performance.
We describe ImageAssist as a decision-support tool, not a diagnostic device. That framing reflects what the software actually does: it informs worklist order, it does not render diagnoses. We do not make FDA clearance claims for the product, and we represent performance metrics as pilot deployment data, not published clinical trial outcomes. The clinical evidence page explains the validation approach in detail.
This is not a limitation to apologize for. It is an accurate description of a tool that is designed to be deployed in a radiologist-in-the-loop model, where the AI contributes to workflow efficiency and the radiologist makes every clinical judgment. That model is the right model for where radiology AI is today.
Questions Worth Asking When Evaluating CDS Tools
When a radiology department evaluates any CDS tool for imaging, the questions that distinguish useful tools from overclaiming ones are:
What specifically does the tool produce as output, and at what point in the workflow does that output appear? What does the radiologist do with that output: review and confirm, measure and document, or simply open the next study in a different sequence? What clinical responsibility changes? What does not change? How is performance measured and by whom?
For worklist triage, the answers are: the tool produces a priority score that updates the worklist position; the output appears before the read begins; the radiologist reads the prioritized study using normal workflow; clinical responsibility does not change; and performance is measured by time-to-read for targeted finding categories against a baseline. That specificity is what makes a CDS tool usable rather than theoretical. If you want to walk through how this applies to your department, reach out to our clinical team.
