Can AI Read My CT Scan? What It Gets Wrong

Your CT produced a stack of images, and what reached you is a PDF with a paragraph called Impression, plus maybe a screenshot you exported from the patient portal at eleven at night. The appointment is a week out. So you ask the obvious thing: can AI just read this for me.

The Short Answer

AI reads the words of a CT report well and the CT images themselves badly, and for CT the gap is wider than most people expect. A CT is a volume the radiologist scrolls through, before and often after iodine contrast, so the one picture you can hand to a chatbot is a slice out of a stack it cannot see. Tested on 99 emergency-department CT cases, GPT-4 with vision named the correct pathology 36.4% of the time (European Radiology, 2025). The text of your report is a different matter: that was written by a radiologist who saw the whole study, and explaining written language is what these models are actually good at.

What AI Is Already Doing To Your CT

Some AI has probably already touched your scan, on the hospital side, before anyone called you. It runs on the full volume, inside the hospital network, and it reports to a radiologist rather than to a patient.

The rib fracture work shows the shape of it. In a study of 243 consecutive chest trauma patients, radiologists working with an AI reader detected rib fractures with 94.2% sensitivity against 69.2% for standard double reading, at 100% specificity (PLoS One, 2025). Counting fractures across twenty-four ribs, slice by slice, is the kind of tedious and well-defined task where a machine beats a tired human.

None of it reaches you. The output of those tools is a flag on a worklist, checked by the person who signs your report, and as RadiologyInfo.org puts it, "AI is designed to support radiologists, not replace them."

Why Your CT Is Hard To Read From A Screenshot

Two things about CT specifically break a general model, and neither applies to a chest X-ray.

The first is the stack. Your radiologist scrolled through the study, forwards and backwards, looking at the same organ across many slices to decide whether something was a vessel in cross section or a nodule. A single exported frame carries none of that, and the finding your report describes may not even be on the slice you saved.

The second is contrast. On a chest CT with contrast, as RadiologyInfo describes it, "the technologist rapidly injects an iodine contrast material into a vein while obtaining CT images." A bright area on a scan taken thirty seconds after that injection means something different from the same brightness on a plain scan, and a screenshot almost never records which one you are looking at. The model will not ask.

CT was the largest group in the emergency-department study behind that 36.4%, 99 of the 216 cases scored for pathology, so this is the modality the model had the most chances at. The same authors recorded hallucinations in 46.8% of its readings. Nothing in the reply separates the two thirds it gets wrong from the third it gets right.

The Three Questions People Bring To A CT Report

A nodule and its millimetres

Nodule sentences are the ones people read twice. What decides the next step is size in millimetres, appearance, and whether it has grown since a prior scan, which is why lung screening CTs are scored with the American College of Radiology's Lung-RADS v2022 system rather than described in free text. Those numbers are already in your report. Getting them out of the report is a reading problem, and our lung nodule guide walks through what each size band usually means.

An incidental finding

CT sees everything in its path, including whatever nobody was looking for. Among 229 asymptomatic adults who had a low-dose non-contrast abdominal CT, 47.2% had at least one incidental finding, kidney findings being the most common, and 15.7% had one the radiologist said needed follow-up (PLoS One, 2025). A cyst mentioned in passing reads like bad news to a patient and like housekeeping to a radiologist.

Whether anything changed

Much of a CT report is a comparison with your last one. The word "stable" carries more of the meaning than any finding does, and it is the word AI reading a lone image cannot produce, because it has never seen your earlier scan. Our guide to what stable means in a radiology report covers the cases where it is genuinely reassuring and the cases where it is not.

What A Reader Built For The Report Does Differently

It starts from the text a radiologist already signed, not from a picture. Our online CT scan reader takes the report you were given and explains what each sentence means: the millimetres, the follow-up interval, the word in the Impression you did not recognize, and the list of things the report rules out, which is usually the part patients skip.

Of the 694 reports ReadingScan has explained since 3 April 2026, 90 were CT studies. Fifteen of those mentioned a nodule and 38 mentioned contrast, which is roughly the mix this page was built for. What it does not do is decide what your scan shows. That decision was made by the radiologist who read the whole stack, and it is already written down in your report.

When Only The Radiologist Will Do

Anything urgent belongs in an emergency department rather than an upload. Anything the report hands to your physician, including the line asking for clinical correlation, is addressed to them and not to you. And where the report itself is ambiguous, the answer is a second human reading, not a second machine.

Frequently Asked Questions

Can AI read the actual images from my CT scan?

Poorly, and CT is the case where this has been measured most directly. On 99 emergency-department CT cases, GPT-4 with vision named the correct pathology 36.4% of the time, and the same study recorded hallucinations in 46.8% of its readings. It will still answer every time you ask.

Why does a single CT image not tell AI enough?

A CT is a volume the radiologist scrolls through, not a picture. The finding described in your report may sit on a slice you never exported, and a screenshot usually does not record whether iodine contrast was in your bloodstream when that slice was taken.

Is AI better at my CT report than at my CT images?

Yes, and the difference is large. Your report is text written by a radiologist who saw the whole study, and turning medical vocabulary into plain English is the task language models are built for. Reading the pictures is a different job with a much weaker track record.

Hospitals already use AI on CT scans. Why can I not use the same thing?

Hospital tools run on the full CT volume inside the hospital network and report to a radiologist who checks them. In one rib fracture study, AI-assisted radiologists reached 94.2% sensitivity against 69.2% for standard double reading. That output never leaves the reading room as an answer to you.

Related Articles

• Lung nodule on a CT scan: what the size and follow-up mean
• Abdominal CT scan report explained
• Incidental finding on a scan: what to do next

---

Disclaimer: This article is for educational purposes only and does not constitute medical advice. Always consult with a qualified healthcare professional for diagnosis and treatment decisions.