Cutting radiology QA turnaround by 60% for a hospital network

A 12-hospital network needed faster chest CT triage without sacrificing diagnostic accuracy. Board-certified radiologists, running through a multi-reader consensus pipeline, got them both.

The challenge

A regional hospital network's radiology department was routing every chest CT triage read through a single on-call radiologist, with a secondary read only for cases flagged as ambiguous. As patient volume grew, average triage turnaround stretched past four hours during peak periods — well past the window where triage actually changes clinical response. The network needed to build a training dataset robust enough to support an AI-assisted triage tool, without introducing the single-reader bias that was already straining their existing workflow.

Our approach

We built a three-reader-plus-adjudication pipeline: every chest CT triage case was independently read by two board-certified radiologists from our network, with a third senior radiologist adjudicating any disagreement. Rather than a binary triage flag, readers scored a structured rubric — urgency, finding location, differential likelihood — so the resulting dataset carried far more signal than a simple "urgent / not urgent" label.

We ran a two-week pilot on 4,000 archived cases before touching live data, calibrating reader agreement against the network's own historical outcomes. Once agreement stabilized above 97%, we scaled to the full active caseload.

The data pipeline

The QA dashboard

Every case flowed through a live dashboard the client's own clinical leadership could see in real time — not just a spreadsheet delivered after the fact.

Results

Across 18 months and roughly 40,000 triaged studies, the consensus-scored dataset supported a triage-assist model that cut average time-to-first-read by 60%, with adjudicated diagnostic accuracy holding at 99.1% against the network's own retrospective outcome data. Just as importantly, the adjudication log gave the client's clinical leadership a defensible audit trail for every disagreement the model was trained on — something their prior single-reader process never produced.

Why it worked

The gains here didn't come from more annotators — they came from routing disagreement to the right expert and keeping a record of why. That's the same pattern behind our broader annotation & labeling program, and it's why healthcare clients consistently see accuracy gains that generic labeling vendors can't match.