Topic 08 · Output & quality · Deep dive
The reading room's hardest numbers: error, workload, and the report that comes too late
A scan's value is only realised when someone reads it correctly and in time. On all three counts the data is sobering. The real-time error rate — about 4% — has barely moved since it was first measured in 1949. The workload behind that number is now one image every three to four seconds. More than half of radiologists report burnout. And in the UK, for the first time on record, not one department could report its images within contracted hours. This is the output side of imaging, measured honestly.
In 1949 the radiologist Leo Henry Garland showed that experienced readers missed important findings on about 30% of abnormal chest films. Three-quarters of a century, and the arrival of CT, MRI and PACS later, the retrospective error rate is essentially the same. This is the defining humility of the field.
1949, and still true
Leo Garland's finding that radiologists miss ~30% of abnormal chest films predates the CT scanner, the MRI, and digital imaging. Every major technology since has improved what we can see — not, measurably, how reliably humans catch it.
Fatigue is measurable
Diagnostic errors cluster in the late afternoon and evening. One ED fracture study found 49% of missed fractures occurred between 4pm and 9pm. Error is partly a function of when in the shift the image is read, not just what it shows.
The litigation tail
Roughly 75% of malpractice claims against radiologists are for diagnostic error, and missed fractures alone account for up to 80% of emergency-department diagnostic errors — a reminder that "output quality" is also a medicolegal exposure.
The error rate hasn't fallen partly because the job has got harder, not easier. Scanners now generate far more images per study than a radiologist a generation ago ever faced, and volumes have outpaced both utilisation and hiring. The single most-cited statistic in modern radiology quantifies the squeeze.
Output quality isn't only accuracy — it's timeliness. The UK's Royal College of Radiologists runs the most complete national audit of this in the world (a 100% response rate from all 159 departments), and its 2024 findings are stark enough to be the clearest window anywhere into a system under strain.
Capital alone won't fix it
The RCR's blunt message to government: doubling scanners doesn't help if no one can read them. 72% of directors said they couldn't even report the output of a new scanner, and 83% couldn't staff one — the bottleneck is people, not machines (echoing Topic 03).
Outsourcing's hidden risk
95% of UK trusts now outsource some reporting. But external teleradiologists often work without prior imaging or clinical context — the RCR flags this as a patient-safety risk, not just a cost. Speed bought at the price of context.
Where AI actually helps
The honest case for AI here (see Topic 04) isn't replacing the read — it's triage and a second set of eyes on perceptual misses, plus admin: ambient transcription, auto-rostering, letter generation. The RCR expects the biggest gains in the administrative sphere.
The number that refuses to improve
If this site has a single most important statistic, it may be this one: the retrospective error rate in radiology is about 30%, and it is essentially unchanged from the day Leo Garland first measured it in 1949. Every intervening advance — computed tomography, magnetic resonance, digital archiving, subspecialisation — has dramatically improved what radiologists can see, without measurably improving how reliably human visual search catches what is there. The real-time error rate, the one that reflects everyday practice, sits around 4%, which sounds reassuringly low until it is multiplied by roughly a billion studies a year to yield something on the order of 40 million errors globally. Neither number is an indictment of radiologists; both are a description of the intrinsic difficulty of the task.
Error is mostly perceptual, and partly about the clock
The structure of those errors matters for what can be done about them. Between 60% and 80% are perceptual — the abnormality was captured in the image but never consciously seen — rather than cognitive misjudgements of a finding that was noticed. Perceptual error is strikingly constant across modality, country and decade, which is what marks it as a property of human visual attention rather than of training or effort. It is also, revealingly, a function of the clock: diagnostic errors cluster in the late afternoon and evening, and one emergency-department study found nearly half of missed fractures occurred between 4pm and 9pm. That temporal signature is the fingerprint of fatigue, and it connects the error rate directly to the workload data.
The workload is the mechanism
The reason the error rate has not fallen is, in part, that the job has intensified faster than any tool has assisted it. A single institution's data captured the shift precisely: between 1999 and 2010 the number of images a radiologist had to interpret per minute rose from under three to over sixteen, because scanners went from producing dozens of images per study to producing hundreds. That is the origin of the field's most-quoted figure — one image every three to four seconds across an eight-hour day — and it is now more than a decade old, predating the most recent surge in volume. More than half of radiologists report burnout on most instruments, some studies far higher, and in the UK the median age at which radiologists leave the profession fell from 55 to 50 in just four years. A workforce losing its most experienced readers mid-career, while volume climbs 8% a year, is the human machinery behind every delayed and every missed report.
Accuracy is necessary; timeliness is the other half
A perfectly accurate report that arrives after a treatment decision has been made has failed the patient as surely as a wrong one. The UK's Royal College of Radiologists, running the most complete national audit in existence with a 100% response rate across all 159 departments, documented in 2024 the first year on record in which not a single department could meet its reporting requirements within contracted hours. More than 434,000 CT and MRI scans went unreported beyond the 28-day target; 16% of patients waited over six weeks for a scan against a 1% target; and £325 million — the equivalent of nearly three thousand consultant salaries — was spent on outsourcing and overtime to paper over the gap. The uncomfortable coda is that this outsourced speed often comes without the prior imaging or clinical context that guards against error, meaning the timeliness fix can quietly reintroduce the accuracy problem. Output quality, in the end, is a single system with two failure modes, and this topic is the honest ledger of both.
On the data. Error and discrepancy rates vary enormously with definition, modality, and study design; "~30% retrospective" and "3–5% real-time" are the most-cited consensus ranges (Berlin, Brady, Bruno et al.), not precise universal constants. The 40-million-errors figure is a simple extrapolation (~4% × ~1 billion studies), not a measured count. The "one image every 3–4 seconds" workload figure is from a single US academic centre (1999–2010) and may not generalise, though its direction is corroborated widely. Burnout prevalence is highly instrument-dependent, hence the wide range. The RCR figures are unusually robust (100% response, single national system) but are UK-specific and reflect a public system under particular strain; they should not be read as global. Figures span 1949–2026 vintages as labelled.