How Often Are CT Scans Wrong? Accuracy and Factors

CT scans get it wrong more often than most people assume, though defining “wrong” matters. A large meta-analysis covering nearly 400,000 CT exams found a total interpretation discrepancy rate of about 8%, with roughly 2.4% of scans containing errors significant enough to potentially change patient care.¹ Those numbers shift dramatically depending on what condition the scan is looking for, who is reading it, what time of day it is read, and whether the scanner itself was set up properly for the job.

What “Wrong” Actually Means in CT Imaging

When researchers measure CT accuracy, they distinguish between several flavors of error. A false negative means the scan missed something that was actually there. A false positive means it flagged something that turned out to be nothing. A “discrepancy” is a mismatch between one radiologist’s reading and a later expert review. These are not the same problem, and they carry different consequences for you. A false negative might delay treatment. A false positive might send you into a cascade of follow-up tests, biopsies, or procedures you never needed.

Across all of diagnostic imaging, average error rates land in the range of 3% to 5%, which translates to roughly 40 million diagnostic errors involving imaging worldwide each year.² That number sounds enormous, but it includes every imaging modality and every severity of mistake. Another way to frame the same data: if only abnormal scans are counted as the denominator, the error rate climbs to about 30%, because abnormal images are inherently harder to read correctly. When you include all scans, normals and abnormals alike, the rate falls to around 3.5% to 4.5%.³ In other words, radiologists rarely tell you something is normal when it is not, but they more frequently struggle with the ambiguous cases where something looks off.

Accuracy Varies Enormously by Condition

There is no single “CT accuracy” number, because the scan’s reliability depends heavily on what it is trying to find. For a common surgical emergency like appendicitis, CT performs well: one five-year review found 137 true positives against only five false negatives and eight false positives, putting sensitivity above 96%.⁴ A systematic review comparing imaging modalities pegged CT’s pooled sensitivity for appendicitis at about 97% and specificity near 96%, making it one of CT’s strongest use cases.⁵ Even so, when clinical suspicion is low, sensitivity can dip to around 88% and specificity to about 82%.⁵

Pulmonary embolism, or PE, is trickier. CT pulmonary angiography has a reported sensitivity around 83% and specificity around 96%, meaning most true clots are caught but a small fraction are missed.⁶ The bigger problem may be overcalling clots. One study had expert chest radiologists re-review initial PE diagnoses and found disagreement in about 26% of cases. The disagreements were heavily concentrated in small, isolated clots: nearly 60% of reported subsegmental PEs were judged negative on expert review, and the most common cause was breathing or beam-hardening artifact mimicking a clot.⁷

Stroke detection is where CT struggles most. A plain head CT has limited sensitivity for early ischemic stroke, particularly when the patient presents with vague symptoms. Among patients visiting the emergency department with vertigo who received a head CT, the scan appeared to miss strokes at a measurable rate: those patients had more than double the risk of being hospitalized for stroke within 30 days compared to matched controls, suggesting the initial CT cleared people who actually had strokes.⁸ CT’s sensitivity for early stroke was measured at 38% to 52% in one study, depending on whether the radiologist knew stroke was suspected.⁹ This is not really a CT failure in the scanner sense; the physics of standard CT simply do not show early ischemic changes as well as MRI does.

The Lung Cancer Screening Problem

Low-dose CT screening for lung cancer in high-risk smokers saves lives, but the false-positive rate is uncomfortably high. In the large National Lung Screening Trial, roughly 27% of scans in the first two rounds came back positive, and the vast majority of those positives turned out to be benign. Across multiple studies, false-positive rates for lung CT screening have ranged from 9% to 50%, with a mean around 20%.¹⁰ Most false positives led to additional imaging. A smaller fraction, about 2.7%, led to invasive procedures like biopsies or bronchoscopy, and among those, roughly one in ten had a complication.¹⁰ The rates did improve with successive screening rounds, dropping to about 17% by the third round, as radiologists could compare new scans against prior ones and recognize stable, non-threatening nodules.

How Fatigue and Shift Timing Affect Readings

Radiologists are human, and their error rates track with exhaustion in predictable ways. A study comparing daytime and overnight body CT readings found that errors were more common at night, and the effect worsened as the shift wore on. In the first half of an overnight assignment, the error rate was about 2.5%. By the second half, it had climbed to 3.7%.¹¹ A separate analysis of 12-hour resident call shifts found that the major discrepancy rate nearly doubled in the final two hours, with an odds ratio of about 1.9 compared to the rest of the shift.¹²

A controlled study testing residents on complex CT cases confirmed the pattern in a different way: fatigued residents detected fewer major findings (an average of about 24 versus 29 when rested) and fewer minor findings as well.¹³ This is not surprising if you think about it the way a pilot would: visual pattern recognition degrades with fatigue, and reading CT images is fundamentally a pattern-recognition task sustained over hours.

Cognitive Biases in Reading Scans

Beyond tiredness, the way radiologists think introduces systematic blind spots. Cognitive biases account for an estimated 74% of all image interpretation errors.¹⁴ These are not random mistakes but predictable mental shortcuts that usually help but sometimes mislead. A radiologist who spots one abnormality may unconsciously stop looking, missing a second finding on the same scan. A radiologist told the patient “probably has pneumonia” may anchor on that diagnosis and miss something else entirely.

Clinical history has a complicated relationship with accuracy. A systematic review found that when radiologists had access to clinical information, somewhere around 38% to 43% of reports changed, and most of those changes improved accuracy.¹⁵ For stroke detection specifically, knowing that stroke was suspected boosted CT sensitivity from 38% to 52% without reducing specificity.⁹ Yet a different study of head CTs for stroke found that detailed clinical history made almost no difference to overall interpretation accuracy.¹⁶ The takeaway is that clinical context generally helps, but it is no magic fix, and it can also prime a radiologist toward the wrong diagnosis.

When the Scanner Itself Is the Problem

Not every error is a human one. Technical factors in how the scan is acquired can make certain diagnoses harder or impossible. Metal implants, dental hardware, and surgical clips produce streak artifacts that obscure nearby anatomy. These artifacts can mask tumors or simulate pathology that does not exist, and they are among the greatest contributors to decreased diagnostic quality in head and neck imaging.¹⁷ Newer reconstruction algorithms can reduce metal artifacts substantially, with one technique cutting average error by 76% compared to standard methods.¹⁸

Scan timing also matters. In CT angiography of the legs, for instance, a patient-tailored scan delay produced significantly higher contrast enhancement throughout the arteries compared to a fixed delay, which showed a drop of nearly 90 Hounsfield units by the time it reached the foot arteries.¹⁹ A poorly timed scan means faint contrast in the vessels you need to see, which can turn a detectable blockage into an invisible one. Patient body habitus, motion during the scan, and even the patient’s cardiac output all influence image quality in ways that have nothing to do with the radiologist’s skill.

The Incidental Finding Trap

A CT scan often finds things it was not looking for. Roughly 20% to 40% of CT examinations contain at least one incidental finding, some abnormality unrelated to the reason for the scan.²⁰ Most of these turn out to be harmless: a small liver cyst, a benign kidney nodule, a thyroid spot that will never cause trouble. But they trigger follow-up. In a study of patients with resected melanoma undergoing routine surveillance CT, 53% had at least one false-positive result or incidental finding. The vast majority, about 88%, turned out benign, but the findings generated 181 additional tests, procedures, and referrals. Fifteen patients underwent unnecessary invasive procedures.²¹

This is not a failure of the CT scan in the usual sense. The scanner correctly identified something that was actually there. The problem is that what it found did not matter clinically, yet the healthcare system responds to it as if it might. Patients with incidental findings but low risk for disease are prone to overdiagnosis and overtreatment, creating what researchers describe as an illusion of benefit while actually conferring harm.²⁰

Resident Readings and the Experience Gap

Many CT scans, especially those done overnight in emergency settings, are first read by radiology residents and later reviewed by attending radiologists. The agreement is generally high. One study of emergency head CTs found 94.6% agreement between resident and consultant readings.²² For body CT, the numbers are somewhat less reassuring: one trauma center reported an overall discrepancy rate of about 6.4% for body CT between resident and attending reads, though only 1% of those were major discrepancies.²³

Experience level matters meaningfully. In a study of nighttime emergency CT interpretations, less experienced residents missed the correct diagnosis 18.3% of the time versus 10.9% for more experienced ones. Body CT interpretations were also more error-prone than neurological CTs, with discrepancy rates of 18.1% versus 9.1%.²⁴ The reassuring part: the clinically important miss rate in that study was 2.3%, and no deaths occurred due to CT misinterpretations during the study period.²⁴

Second Opinions and Double Reading

Getting a second radiologist to review the same scan catches errors at a meaningful rate. A systematic review of double reading in diagnostic radiology found discrepancy rates between 0.4% and 22%, depending on the setting, with sub-specialist review generally leading to higher rates of changed reports.²⁵ In neuro-oncology, where the stakes are high and images are complex, second-opinion interpretations disagreed with the original reading 19% of the time. In cases with enough follow-up to verify who was right, the second opinion was correct 100% of the time, and the report differences would have changed patient management in about 15% of all cases reviewed.²⁶

If you are facing a serious diagnosis based on CT findings, requesting a subspecialist second opinion on the imaging is one of the most practical steps you can take. Most academic medical centers offer this service, and for cancer staging or surgical planning in particular, the evidence suggests it can change the course of care in a meaningful fraction of cases.

What AI Is and Is Not Doing Yet

Artificial intelligence tools are entering radiology workflows, mostly as assistants that flag potential abnormalities for human review. Across studies, AI applied to CT has shown improved detection rates of up to about 20% for specific tasks like lung nodule detection and intracranial hemorrhage triage, along with reduced turnaround times in real-world settings.²⁷ For coronary CT angiography, one automated system reached strong agreement with expert readers and had a very high negative predictive value of 98%, meaning it was excellent at confirming that a person did not have severe coronary artery disease. Its positive predictive value was only 39%, though, meaning it flagged many cases as severely diseased that were not.²⁸

That pattern, strong at ruling things out and weaker at ruling things in, is common with current AI tools. They work best as a safety net that catches things a tired human might miss, rather than as a standalone diagnostician. The technology is improving, and deep learning models for PE detection have reached pooled sensitivities around 88% and specificities around 86%.²⁹ But no AI system has replaced the need for a human radiologist to make the final call.

CT Errors and Malpractice Claims

CT-related misdiagnosis does show up in legal data, though it accounts for fewer claims than you might expect given CT’s volume. In an analysis of medicolegal closed claims, CT had a risk ratio of 1.9 for generating a malpractice claim compared to the baseline across modalities. Mammography (risk ratio 4.0) and MRI (risk ratio 3.4) were both higher risk per exam.³⁰ Across all modalities, the dominant error type was misdiagnosis, meaning a delay in or failure to correctly read the imaging, which accounted for 62% of claims.³⁰ A retrospective analysis of radiology malpractice cases in China found a similar pattern, with diagnostic errors making up about 80% of primary medical errors and missed diagnoses specifically accounting for 68% of those cases.³¹

The legal data underscores a point the clinical literature also makes: the most dangerous CT error is not finding something that was never there. It is failing to see something that was there all along, often because it was subtle, the radiologist was fatigued, or the clinical question sent them looking in the wrong direction. Systems that improve communication of unexpected findings are being developed to close that gap. One electronic health record program designed to track unexpected nonemergent findings flagged potential neoplasms in about 11% of submitted examinations, ensuring those findings were not lost in the shuffle between radiologist and treating physician.³²

How CT Compares to Other Imaging

For the conditions where CT is most often used, it generally matches or outperforms ultrasound and performs comparably to MRI. In acute abdominal pain during pregnancy, where MRI is often preferred to avoid radiation, both CT and MRI achieved 88% diagnostic accuracy with no statistically significant difference.³³ For appendicitis, CT and MRI sit in the same tier with sensitivities above 96%, while ultrasound lags behind with pooled sensitivities in the low to mid 80s and a non-diagnostic scan rate that can reach 36%.³⁴

CT’s real advantage over MRI is speed, availability, and cost. An MRI takes longer, costs more, and is not always available in the middle of the night. CT’s real disadvantage is radiation exposure and lower soft-tissue contrast, which makes it weaker for conditions like early stroke or subtle brain tumors where MRI shines. The choice between imaging modalities is usually not about which has better accuracy in a vacuum, but which one gives the best answer for a specific clinical question at the moment it needs answering.