What AI Can—and Cannot—Explain About Medical Imaging Reports

Artificial intelligence can turn a dense block of medical terminology into language that feels more approachable. For someone reading a CT, MRI, ultrasound, or X-ray report, that capability may reduce the effort required to identify sections, look up unfamiliar terms, and prepare questions.

But a fluent explanation can also create false confidence. A system may produce clear, authoritative-sounding text even when it lacks essential clinical context or misunderstands the source material. The safest way to use patient-facing AI is to define a narrow role: help people navigate information while preserving the authority of the radiologist and treating clinician.

What AI may help explain

Report structure

AI can identify common sections such as the clinical indication, technique, comparison, findings, and impression. It can explain why each section exists and direct attention to recommendations or comparison language.

This is a navigation task. It helps a patient understand how the document is organized without deciding what the findings mean for that individual.

Medical terminology

AI can provide general definitions for terms commonly used in radiology. For example, it may explain that an artifact is a feature caused by the imaging process rather than anatomy, or that incidental describes something discovered unexpectedly and not necessarily related to the reason for the examination.

A responsible definition should retain important qualifiers. It should not silently convert “may represent” into “is,” or “cannot exclude” into a confirmed diagnosis.

Plain-language restatement

AI may restate a sentence in simpler language while keeping the original report visible. This can make it easier to locate the source of an explanation and compare the simplified wording with the radiologist's text.

Good restatement preserves:

• Negation, such as “no evidence of”;
• Uncertainty, such as “possibly” or “indeterminate”;
• Comparisons, such as “stable since the prior examination”;
• Measurements and units;
• Laterality, such as left versus right;
• Recommendations and time frames.

Questions for a clinical appointment

AI can help convert confusion into questions. Examples include asking whether a finding is new, how it relates to symptoms, why follow-up was recommended, or which professional should discuss the result.

This use can strengthen the patient-clinician conversation because it organizes questions rather than attempting to answer them conclusively.

What AI cannot safely determine on its own

A patient-specific diagnosis

A radiology report is only one component of a clinical assessment. Diagnosis may depend on history, physical examination, laboratory results, pathology, medications, previous imaging, family history, and specialist judgment. A general-purpose AI system usually does not have all of this information, and having more data does not automatically make it clinically qualified.

An AI explanation should therefore avoid statements such as “you have,” “this proves,” or “this rules out” unless it is accurately quoting the report—and even then, it should distinguish a quoted radiology conclusion from its own output.

Whether a finding is dangerous

Risk cannot reliably be inferred from one phrase. Size, location, appearance, change over time, symptoms, age, and other factors can alter clinical significance. A reassuring-sounding explanation can delay care; an alarming explanation can create unnecessary distress.

Treatment or follow-up decisions

AI should not independently tell a patient to begin, stop, or change treatment. Nor should it replace a radiologist's or clinician's follow-up recommendation with a different interval or test. If a report contains a recommendation, the tool can surface it and encourage the patient to discuss it with the responsible professional.

The meaning of the images themselves

Explaining report text and interpreting original medical images are different tasks. Image interpretation requires validated systems, appropriate image quality, relevant views or sequences, and clinical expertise. A text explanation must not imply that the images were independently reviewed when only the report was processed.

Urgency from incomplete information

Some report language can indicate that timely communication or follow-up is important, but urgency also depends on symptoms and clinical context. Patient-facing tools should provide clear escalation guidance for emergencies without pretending to triage every situation.

Fluency is not evidence of correctness

The World Health Organization has warned that large language models can produce responses that appear authoritative and plausible while containing serious errors. WHO recommends transparency, expert supervision, rigorous evaluation, protection of autonomy, and accountability for AI used in health (WHO calls for safe and ethical AI for health).

This problem is especially important in radiology because a small change in wording can reverse meaning. Missing the word “no,” changing left to right, dropping a measurement unit, or overstating an uncertain finding can materially alter the message.

Users should be able to see the original report beside the explanation and identify which source text supports each summary. When a tool is uncertain, it should say so rather than filling gaps with a confident guess.

Transparency should match the intended audience

The FDA, Health Canada, and the UK Medicines and Healthcare products Regulatory Agency describe transparency as communicating relevant information about intended use, development, performance, and logic to the appropriate audience. Their joint principles emphasize the human-AI team, clear essential information, user context, and human-centered design (Transparency for Machine Learning-Enabled Medical Devices).

For a patient-facing explanation tool, meaningful transparency includes:

• Whether the system reads report text, images, or both;
• What the tool is intended to do;
• What it is not intended to do;
• Whether it is a regulated medical device for a particular use;
• Known limitations and common failure modes;
• How personal health information is handled;
• When a clinician should review the result;
• How users can report an error or request support.

A long terms-of-service document is not enough. Critical limitations should appear at the moment they matter.

A practical boundary: education before interpretation

A safer patient-facing experience can follow a progression:

1. Orient: identify the report sections and exam type.
2. Define: explain terminology in general language.
3. Restate: simplify sentences without removing uncertainty.
4. Trace: connect every explanation to the original text.
5. Question: help the patient prepare questions for a professional.
6. Escalate: clearly identify situations requiring clinical or urgent attention.

The system should stop before it independently diagnoses, predicts an individual's outcome, or recommends treatment.

How patients can use AI explanations more safely

• Remove unnecessary identifying information before using a service, unless its privacy protections and intended use are clear.
• Confirm that the correct report and examination date were used.
• Compare the explanation with the original report.
• Check whether uncertainty, negation, measurements, and left/right orientation were preserved.
• Do not use an AI answer to cancel follow-up or change treatment.
• Bring unresolved questions to the ordering clinician or radiologist.
• Seek urgent care based on serious symptoms, not reassurance from an automated explanation.

The right success measure

The goal of patient-facing imaging AI should not be to produce the most definitive-sounding answer. It should be to improve understanding without increasing unsafe self-interpretation.

Useful measures might include whether patients can identify the purpose of the examination, understand which statements are uncertain, recognize follow-up recommendations, and formulate better questions. Accuracy and safety should be evaluated across different populations, reading levels, report types, and clinical contexts.

AI can make a radiology report easier to approach. It cannot turn that report into a complete clinical assessment, and it should never obscure that boundary.

References

• World Health Organization. Ethics and Governance of Artificial Intelligence for Health.
• World Health Organization. WHO Calls for Safe and Ethical AI for Health.
• U.S. Food and Drug Administration. Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles.
• RadiologyInfo.org. How to Read Your Radiology Report.