The Future Radiologist May Not Read the Whole Study Alone

Imagine a patient having a CT scan to see whether a tumour in their lung has changed. The radiologist needs to measure it, compare it with an earlier scan, check for new findings, and explain what the results mean for the patient. Each step contributes to one report, but each asks a different question.

Today, the work usually follows a familiar pattern: one study, one radiologist, one report. The whole study arrives as a single assignment. Even AI is mostly fitted into this structure: it may help with a measurement or highlight a finding, while the radiologist manages the rest. Looking inside that assignment reveals several diagnostic tasks. Some require careful measurement, some need a search through earlier images, and some need a specialist’s judgment.

The next step could be to organize support around those tasks. Instead of sending a whole study to one person and adding tools around them, the system could match each question with the expertise suited to it:

These are possible routes, not a checklist for every scan. A measurement may need only a quick check. An unclear finding may need a colleague’s opinion. Many studies will need no additional doctor.

The primary radiologist keeps ownership of the study. Other people and tools can contribute to defined tasks, while the radiologist checks their work and takes responsibility for the final report.

Some tasks can be prepared before the radiologist begins reading. Measure a lesion on a follow-up CT. Its size matters because the referring team needs to understand whether the disease has changed. Consistently measuring the same lesion across scans takes time, especially when several findings need tracking.

In a study of 55 patients with melanoma, three radiologists outlined lesions on follow-up CT scans manually and with AI assistance. Outlining a lesion allows you to measure its volume. With AI, time on this task fell by 33%, from 336 to 222 seconds, while outline quality remained similar.

Finding the right images for comparison is another task. In a separate study, software helped nine radiologists locate the same finding on earlier CT and MRI scans. Their average assessment time fell from 107 to 65 seconds. The radiologists still examined and measured the finding themselves; the tool helped them reach the images they needed.

Both studies tested one defined step, not the assessment of a whole scan, and that is the point. The tool handles part of the process, and the radiologist uses the result to answer the larger clinical question. When the question is what a finding means, you may need a different kind of help.

A radiologist may encounter a finding that is difficult to interpret or one whose meaning could affect treatment. Two forms of help are possible. A subspecialist brings experience with a particular image type or disease. Another radiologist can provide an independent second opinion when an interpretation remains uncertain. The roles may overlap, but the question sent to a colleague should be clear.

Research on second readings shows why access to specialist expertise matters. At one cancer center, specialists reviewed CT and MRI scans of patients referred from other hospitals. In 42 of 283 cases, differences between the original and second reports could have changed treatment or disease staging.

It also matters when a radiologist can get help from a colleague. Specialists reviewed 464 imaging exams of children performed during on-call hours. They found errors in about one in five initial reports. Around half of those errors could have affected patient care.

Both studies reviewed existing reports. The cancer center scans had been selected for another review, so the figure shows the value of specialist input in those cases, not how often every scan needs it. The pediatric study did not test a system for choosing which exams to send. These are situations where access to expertise can matter, including complex oncology cases and on-call hours.

The system needs a way to recognize when a task calls for additional expertise. Sometimes the radiologist knows a finding is unclear, and a focused consultation would help. In other cases, the type of study or a rare condition may point to a subspecialist from the start. A future system might also suggest a review when an AI result differs from the radiologist’s assessment.

A suggestion is only useful if it reaches someone who can act on it. Too many alerts could add work and delay cases that need attention. Too few could leave important disagreements unseen.

Selecting the right tasks and routing them to the right expertise is therefore a core coordination problem. The aim is to bring in human or AI support where it is needed, while the primary radiologist retains the clinical context and final responsibility for the study. This coordination needs testing alongside the tools that produce the signals.

Quality assurance studies offer examples of how systems flag cases for another look, although they examine reports after completion rather than consultation before final sign-off.

In one study of more than 111,000 high-acuity CT scans, review was triggered when the report was negative, AI was positive, and the radiologist had not opened the AI output. Only 46 scans met these conditions; review confirmed a missed finding in 26.

In another study, AI flagged 5,289 possible discrepancies between 25,104 chest X-rays and their reports. An external radiologist selected 47 cases for further review, and hospital radiologists confirmed important missed findings in 35.

Both checked reports after they had been completed, using different rules. They cannot tell us when to request a second opinion while a radiologist is still working on a scan. They do show why a person needs to review AI alerts: most of the alerts in the second study did not lead to a confirmed missed finding.

Once a question goes to another participant, the answer has to come back, and every handoff is a place where information can be lost or arrive too late. A useful request needs the relevant images, enough information about the patient, and a precise question: for example, whether a particular finding changes the assessment of the disease. The radiologist also needs to know when a reply has arrived and whether it answers that question.

At one medical center, radiologists received an electronic alert when a colleague disagreed with their report. They had to confirm that they had received it. In 64 of 114 such cases, the second opinion changed the patient’s care plan. The original report had already been completed. The study shows why a second opinion needs to reach the radiologist responsible for that report.

When software and other doctors help with parts of a study, someone needs to bring their results together. A measurement may be accurate and a specialist’s opinion may be useful, but the radiologist still needs to decide what they mean for the patient. The radiologist responsible for the study makes the final assessment.

Return to the patient’s CT scan. Software could prepare measurements and bring up the earlier images. The radiologist would check that work and review the full study. If a new finding required specialist input, they could send a focused question, assess the reply, and write one clear report for the referring doctor.

This changes the basic unit of diagnostic work: from routing an entire study to one reader to routing individual diagnostic tasks to the expertise best suited to them.

The future of radiology may not be about distributing studies between radiologists. It may be about assembling the right expertise around each study.