Imaging center ops dashboards tend to collect everything. Volume by day, scanner utilization, technologist productivity, per-payer reimbursement tables. Most of it sits in an RIS report that someone exports to Excel once a week and reads when something already went wrong.
The metrics that actually tell you where throughput is breaking down are a much shorter list. We track five of them as the core signal set when we're looking at how a pre-reading layer integrates into an imaging center's daily operations. Not because the others don't matter, but because these five expose the specific friction points that a workflow tool can actually move.
1. Study Volume by Modality, Trended Daily Not Weekly
Total study count is the most basic operational number, but the granularity matters more than most centers recognize. Weekly or monthly aggregates hide the Monday accumulation pattern that affects nearly every center: the weekend queue hits the readers in a block first thing Monday morning, and TAT for Monday studies is structurally worse than Tuesday's. If you only look at weekly averages, that Monday spike is invisible.
Breaking volume by modality (CT, MRI, X-ray, US) on a daily basis gives you a very different picture than aggregate volume. A 15% increase in CT volume in a week where MRI held flat and X-ray dropped is actionable: CT pre-reads are more time-intensive for the radiologist, and the queue composition matters for predicting turnaround. Centers that track blended volume as a single number miss those composition shifts until they show up as TAT complaints from referring physicians.
Daily modality volume also sets the denominator for the next metric.
2. Pre-Read Coverage Rate
If you're running an AI pre-reading layer, the first operational question is simply: what fraction of studies that hit the worklist have a draft ready when the radiologist opens them? We call this coverage rate. It is not a quality metric. It's a plumbing metric.
Coverage rate below roughly 90% means there's a gap somewhere in the routing logic, the DICOM send trigger, or the modality filter configuration. One imaging center we worked with in 2024 had a coverage rate sitting at around 71% and couldn't identify why. The gap turned out to be a single mammography unit that was sending studies via a different AE title than the others. Every mammo study was bypassing the pre-read queue entirely. That's not an AI problem. It's an integration configuration problem, and it only becomes visible if you track coverage rate explicitly.
Coverage rate should be tracked per modality and per sending device. A blended 95% coverage rate can hide a single scanner or modality that's consistently at 60%.
3. Draft Acceptance Rate by Modality
Draft acceptance rate measures what percentage of AI-generated drafts a radiologist accepts as the basis for their final report, versus dictating from scratch. This is the single most informative metric for evaluating whether the pre-reading output is actually useful to the readers in your environment.
We're not saying high acceptance rate equals high accuracy or that a low acceptance rate means the AI is wrong. What it actually measures is fit: does the draft language match how your radiologists document findings, for your patient population, for the specific mix of studies you see? A draft written for a high-acuity hospital trauma population will look meaningfully different from what a community imaging center reads all day, even for the same modality.
The modality breakdown is important here. Acceptance rate for chest X-ray in a stable outpatient population often runs considerably higher than acceptance rate for MRI of the brain in a mixed-indication population. Treating them as a single number obscures that, and leads to policy decisions (like "should we use drafts for MRI?") based on blended data that doesn't reflect the actual distribution.
Track acceptance rate at the individual radiologist level as well. Not for individual accountability, but because systematic outliers usually indicate a workflow friction point: a radiologist who never accepts drafts may be working in a dictation configuration where inserting the draft requires an extra click that others don't have.
4. Report Turnaround Time by Modality, Segmented by Day Part
TAT is the metric most imaging centers already track. The problem is usually how they segment it. Overall mean TAT for all studies across a day or week is nearly useless for operational decisions. The signals are in the segments.
Modality segmentation is straightforward: CT and MRI reads take longer than plain film, and bundling them together tells you nothing about where the bottleneck is.
Day-part segmentation is less common but more valuable. TAT for studies arriving between 8am and 10am (the morning surge) is structurally different from studies arriving between 2pm and 4pm. Understanding that pattern tells you whether you have a queue-at-open problem (studies accumulating overnight and on the weekend) or a capacity problem during a specific window.
When a pre-reading layer is in place, you want to track TAT for studies with a draft available versus studies where the radiologist opened with no draft. The difference between those two cohorts, controlled by modality and day part, is the cleanest signal for the operational value of the pre-read. If there's no difference, the drafts aren't being used. If the difference is large in one modality and small in another, that guides where to invest in improving draft quality or radiologist workflow adoption.
5. Urgent Flag Rate and Escalation Acknowledgment Time
The fifth metric is less about throughput volume and more about the quality signal that lives inside throughput. What percentage of studies receive an urgent flag, either from the AI pre-read or from the radiologist on opening? And for flagged studies, how quickly does the ordering physician or facility receive notification?
Urgent flag rate as a percentage of total volume should be roughly stable for a given center and population mix. A sudden spike may indicate a real population shift or a calibration issue with the flag logic. A sudden drop may indicate under-flagging, which is a patient safety concern.
Escalation acknowledgment time is the time between the urgent flag being set and the downstream acknowledgment that the referring provider received and acted on the alert. This is an area where imaging centers frequently have no structured measurement at all. A study flagged for possible pulmonary embolism at 2pm that is not acknowledged until 6pm is a workflow failure that aggregate TAT metrics won't catch.
We've seen centers where the AI pre-read caught a pattern suggesting intracranial hemorrhage, the radiologist reviewed and confirmed urgency within minutes, but the communication pathway to the ordering physician was a fax queue that wasn't checked for two hours. No TAT number captures that. Escalation acknowledgment time does.
Using These Five Together
Each of these metrics answers a different question, and the combination is more diagnostic than any single number. Volume tells you what hit the system. Coverage rate tells you how much of that volume entered the pre-reading workflow. Acceptance rate tells you whether the output was useful. TAT segmentation tells you where time is being consumed. Urgent flag rate and acknowledgment time tell you whether the critical minority of studies got the fastest path.
When we're configuring a new deployment, we spend the first two weeks establishing baselines across all five before drawing any conclusions. A center that runs 180 studies per day across three modalities and has a blended TAT of 90 minutes will look very different from one at the same volume where 40% of studies are CT, the Monday queue is 60% larger than the rest of the week, and urgent flag acknowledgment runs on a fax system. The same tool deployed in both environments will behave differently, and measuring both environments with the same blended metrics will give you false equivalence.
None of these metrics require new software to track. Most are derivable from existing RIS exports and PACS event logs. The work is mostly in the segmentation and the cadence: daily beats weekly, modality-level beats blended, and cohort comparison (draft-available versus not) beats single-number averages.