Radiologists vs. AI: The Machine Sorts the Worklist, but Humans Still Read the Scans
Of the medical AI tools cleared by the FDA, 76% are in radiology. They already triage CT scans, draft reports and cut the mammography reading workload almost in half. Yet demand for radiologists keeps climbing, and Israel is short about 300 of them. What is shifting to the machine, what stays with doctors, and what to do now.
The profession and what's changing
Radiologists are physicians who specialize in interpreting medical imaging: X-rays, CT, MRI, ultrasound and mammography. The scans themselves are usually performed by radiographers (radiology technologists). The radiologist then receives the images, looks for findings, compares them with previous studies and writes the report that treating physicians rely on. Some radiologists also perform image-guided procedures such as biopsies and catheterizations.
For years, this was the profession first in line to be replaced. About a decade ago, AI pioneer Geoffrey Hinton declared that we should stop training radiologists because machines would replace them within five to ten years. It didn't happen. What is actually happening is a new division of labor: the machine sorts, flags and drafts, while the doctor reviews, corrects and signs off. At the same time, the workload isn't shrinking but growing.
What AI already does today
Radiology has more cleared tools than any other field of medicine. According to the U.S. Food and Drug Administration's (FDA) list, 1,451 AI-enabled medical devices had been authorized by the end of 2025, and 1,104 of them, or 76%, were in radiology. In 2025 alone, the field accounted for 75% of authorizations. The tools mainly do three things:
- Triage by urgency. In early 2026, Israeli company Aidoc received FDA clearance for a system built on CARE, a foundation model it developed, which scans abdominal CT studies to move urgent findings up the worklist. According to the company, the system is expected to expand to all CT and X-ray workflows within 18 months, and Aidoc is also developing automated drafting of preliminary reports.
- A second reader in mammography. In Sweden's MASAI trial, a randomized controlled trial whose results were published in The Lancet journals, ScreenPoint's Transpara system was integrated into breast cancer screening. According to the company's summary, adding AI increased the cancer detection rate by 29% and reduced the reading workload by 44% compared with the standard approach, in which every exam is read by two doctors. The final results, published in January 2026, showed 12% fewer interval cancers, those detected between screening rounds, a reduction the trial deemed "non-inferior", meaning at least as good as the existing method, with no increase in false positives.
- Draft reports. At Northwestern Medicine in Chicago, researchers developed a generative model that drafts reports for X-rays and flags life-threatening findings. In a study published in JAMA Network Open, the system was deployed across 12 hospitals for five months on nearly 12,000 images. Report-writing efficiency rose by an average of 15.5%, and by as much as 40% for some radiologists, with no loss of accuracy.
None of this has translated into layoffs. In the U.S., a record 1,208 radiology residency positions were offered in 2025, up 4% from 2024, and the share of unfilled positions in the field hit an all-time high. In Israel, the shortage is more acute. According to Dr. Arnon Makori, head of imaging at Assuta, roughly 600 to 650 radiologists practice in the country, about six per 100,000 residents, compared with a European average of 12, leaving a shortfall of about 300. He cites lengthy training and too few residency slots. New imaging machines, some of which use AI to shorten scan times, actually increase the workload, he says. Screening programs added to Israel's state-funded health basket, such as low-dose CT to detect lung cancer in former smokers aged 65 to 74, are piling on more.
What remains human
The full picture. Most tools are narrow: each is cleared for a defined task, such as brain hemorrhage, pulmonary embolism or a breast mass. The radiologist is responsible for everything in the scan, including incidental findings no one was looking for. Even Aidoc's broader system is a triage tool that moves cases up the queue, not a full read.
Clinical context. The same shadow on the lung means one thing in a patient recovering from surgery, another in a woman undergoing chemotherapy and a third in a 70-year-old smoker. The radiologist reads the referral, compares prior studies, consults with the treating physician and recommends the next step. That is judgment based on information that isn't in the image.
Accountability. A radiology report is a medical document signed by a physician. In the studies described here, AI worked alongside doctors, not instead of them: in Sweden it reduced the need for the second read, not the first; at Northwestern, radiologists completed and approved the reports based on the draft.
Overseeing the machine. An algorithm that performed well at one hospital can stumble on a different scanner or in a different population. Aidoc itself lists continuous performance monitoring as part of its system. Someone with clinical expertise has to decide which alerts are reliable and when the tool is drifting.
The hands. Interventional radiology, in which doctors perform biopsies, drainages and catheterizations, is hands-on work at the patient's side, and AI serves there as an assistive tool at most.
Where people in the field are heading
- Subspecialties. Breast imaging, neuroradiology and interventional radiology, where cases are more complex or involve procedures, are the areas where a physician's judgment and hands matter most.
- Clinical AI leads in hospitals. The people who select tools, test them on local data before deployment and track their performance. It's a hybrid role that blends physician, IT manager and quality officer.
- The medical imaging industry. Companies like Aidoc and ScreenPoint need physicians for data labeling, clinical validation and regulatory submissions, alongside biomedical engineers and data scientists.
- The new version of the same job. A workday in which the algorithm has already sorted the worklist and proposed a draft, while the doctor reviews more studies, focuses on the outliers and spends more time communicating with the care team.
What to do now
If you already work in the field
- Learn to ask the right questions about every tool: what data it was trained on, what its false-positive rate is and which population it was tested on.
- Join the deployment committees at your hospital or imaging center, rather than leaving the decision to procurement.
- Use report-drafting tools, but check every sentence: your signature is still on it.
- Consider a subspecialty where clinical judgment or hands-on procedures are central.
If you're thinking of entering
- The path runs through medical school and a radiology residency. Training is long and slots are limited, but Israel's shortage points to continued demand in the coming years.
- The profession suits people who enjoy precise visual work and sustained concentration but can also communicate with a team. The workload is heavy, and so far AI hasn't lightened it.
- If you're interested in imaging but not in medicine, consider a degree in medical imaging, which trains radiographers, or biomedical engineering and data science roles at imaging companies.