What the study found
Researchers at the Walton Centre NHS Foundation Trust in Liverpool, a UK tertiary neuroscience center, reviewed 580 consecutive formal neuroradiology second-opinion reports generated between April 2024 and March 2025 — 487 MRI studies (84%) and 93 CT studies (16%). Each second-opinion report was compared against the original interpretation and classified as agreement, minor discrepancy, or major discrepancy, then coded by likely error type. The analysis, led by Dr. Nasr Abdelsalam and colleagues, was covered by Radiology Business after being published in Clinical Radiology.
The topline number: 245 of 580 studies (42.2%) — roughly 2 in 5 — showed some discrepancy between the original report and the subspecialist's second read. Of those, 124 cases (21.4% of all second opinions) were major discrepancies, and 121 (20.9%) were minor. Put simply, just over 1 in 5 formal neuroradiology second opinions surfaced a read significant enough to matter clinically — not a rounding error, a routine outcome.
The consequences of a major discrepancy were substantial. Among the 124 major-discrepancy cases, 123 (99.2%) involved a changed diagnosis, 116 (93.5%) led to a change in treatment, and 85 (68.5%) prompted further investigation. Vascular imaging findings accounted for nearly half of all major discrepancies — 61 of 124 cases (49.2%) — making it the single largest category of high-stakes disagreement.
Why the misses happen
The researchers also coded the likely cause of each discrepant read. The pattern points squarely at cognitive error rather than technical failure: faulty reasoning accounted for 109 of 245 discrepant cases (44.5%), complacency for 54 (22%), underreading for 37 (15.1%), and satisfaction of search — stopping the review once one abnormality is found — for 28 (11.4%). As the authors put it, neuroradiology is "particularly susceptible to diagnostic error due to the complexity of neuroanatomy," and they noted the field has lacked "contemporary real-world data evaluating the clinical impact of formal second-opinion reporting."
None of these error types are unique to any one radiologist or center — they are the well-documented failure modes of single-reader interpretation under time pressure, and the reason double-reading and structured second opinions exist as a quality-assurance concept in the first place.
| Error type | Share of discrepant cases | Cases |
|---|---|---|
| Faulty reasoning | 44.5% | 109 of 245 |
| Complacency | 22.0% | 54 of 245 |
| Underreading | 15.1% | 37 of 245 |
| Satisfaction of search | 11.4% | 28 of 245 |
The catch: almost nobody gets this kind of second opinion
Here is the part the topline number obscures. A formal, subspecialist second-opinion service like the one studied at the Walton Centre is not a routine part of most patients' care — it exists for referrals into a tertiary center, complex oncology cases, or curbside consults a referring physician happens to request. The vast majority of CT and MRI studies performed every day get exactly one read, by one radiologist, once.
That is the gap this data quietly exposes. If a dedicated subspecialist second look changes the clinical picture in roughly 1 out of every 5 cases it's given, the studies that never reach a second reader are not exempt from that same error rate — they simply never get the chance to have it caught. Ad hoc, curbside second opinions reach a small, often self-selected slice of studies: the ones a referring clinician already suspects are wrong, or a patient pays out of pocket to have reviewed. Everyone else's single read stands, discrepancy or not.
Where a structured, always-on second read fits
The evidence argues for making a second read a standard part of every study's workflow, not a discretionary extra. That is the structural gap AI-assisted CT reporting is built to close: an AI system drafts a comprehensive, structured report on every case, an in-house radiologist reviews that draft before it goes out, and the finding lands on the referring physician's desk ready-to-sign by the client's own reading radiologist. It doesn't replace the value of subspecialist consultation for genuinely hard cases — it extends the same "someone else looks before it's final" mechanism to the routine studies that a formal second-opinion service was never built to reach at scale.
Frequently asked questions
How often do radiology second opinions find a discrepancy?
In a 2026 UK study of 580 consecutive neuroradiology second-opinion reports, 42.2% showed a discrepancy between the original report and the subspecialist's interpretation. Just over 1 in 5 of all second opinions (21.4%) involved a major discrepancy — one significant enough to affect patient management.
What is the difference between a major and minor radiology discrepancy?
A minor discrepancy is a difference in interpretation that would not meaningfully change how a patient is treated. A major discrepancy is one where the second reader's finding would change management. In the UK neuroradiology study, major discrepancies led to a change in diagnosis in 99.2% of cases and a change in treatment in 93.5% of cases.
What causes most radiology second-opinion discrepancies?
The study attributed discrepant reads mainly to cognitive and interpretive error rather than image quality: faulty reasoning accounted for 44.5% of discrepant cases, complacency 22%, underreading 15.1%, and satisfaction of search — stopping the review after finding one abnormality — 11.4%. Vascular findings were involved in nearly half of all major discrepancies.
Do most patients actually get a second radiology read?
No. Formal second-opinion services like the one studied are typically reserved for complex referrals to a tertiary center, not routine imaging. Most CT and MRI studies get a single read. A structured second read on every study — an AI-generated draft plus a radiologist review before the report reaches the referring physician — extends the same discrepancy-catching mechanism to studies that would otherwise never get one.
Source: Abdelsalam N, Mills S, et al., Walton Centre NHS Foundation Trust, published in Clinical Radiology (2026), as reported by Radiology Business. Figures are rounded as reported.