Radiologists miss lung nodules in roughly 30% of abnormal CT scans, a failure rate that directly harms patients. One study found that more than half of patients later diagnosed with lung cancer had visible nodules in their scans from the previous year that went undetected. New research from the New York Institute of Technology is now examining why these misses happen and how AI tools can be designed to reduce them without introducing new risks.
Robert G. Alexander, Ph.D., a cognitive neuroscientist and assistant professor of psychology and counseling, leads the Human Factors and Neuroscience (HFAN) research group. His team recently published a review in the American Journal of Roentgenology that categorizes the causes of missed nodules into three groups: nodule characteristics like small size and low density, technical factors such as poor image quality, and reader limitations including fatigue, distraction, and over-reliance on AI tools.
"If a radiologist misses an abnormality, you're not treating the patient. That immediately can lead to potential patient harm. So, we care a lot about misses," Alexander said. "And in most instances, we don't really know exactly why a miss happens."
When AI tools help - and when they don't
The published evidence on AI-assisted detection is split. Some studies show AI tools boost radiologists' nodule detection by 24% and reduce interpretation time, particularly when the tools work as assistants alongside the radiologist. Other studies found no clear improvement. The same nodule types that radiologists most often overlook - those with lower density and poorly defined margins - are also the ones where AI tools prove least sensitive.
False positives add another layer of risk. AI tools sometimes flag areas that are not clinically concerning. Radiologists can become biased toward agreeing with the tool's output, which leads to unnecessary follow-up procedures and patient anxiety. "Some of the tools are quite good and useful, and it's more a question of useful in what way, and what is the right way to approach them," Alexander said.
Testing how cues shape radiologist performance
Alexander's lab, funded by a National Institutes of Health grant, is now one year into a four-year project measuring how radiologists respond to different types of AI-generated cues. The team is comparing general prompts ("find the nodule") against precise prompts that include descriptive details ("find the lobulated nodule"). About 50 radiologists have participated so far.
The researchers expect precise cues will help radiologists locate concerning areas faster and with fewer eye movements. But they are also testing an unintended consequence: narrowly focused attention might cause radiologists to miss other abnormalities elsewhere in the image. This project is the first to quantify how these cues affect radiologists' eye movements as they interpret CT scans, tracking where they look and how they scroll through image stacks. The findings will inform how future AI systems are designed for clinical workflows, an area closely tied to developments in AI for Healthcare.
Undergraduate research drives the work forward
Talia Lilikakis, a first-year osteopathic medical student in the B.S./D.O. program, joined Alexander's group after taking his introductory psychology course as a freshman. She became the second author on the AJR review and presented its findings at the 2026 Annual Meeting of the Society of Thoracic Radiology, where she won the award for Best Student Oral Scientific Presentation.
Lilikakis also helped conduct additional research at the conference, examining how experienced radiologists train newcomers to interpret medical images. She plans to focus on neurology and continue integrating research into her career. "As AI becomes more and more prevalent in everything that we do, and especially in medicine, I feel like it has the potential to be very helpful if it's used correctly. And so, while it's growing, I think proper research needs to be done now," she said.
Why this matters for science and research professionals
The HFAN lab's approach combines cognitive science, eye-tracking methodology, and AI evaluation to tackle a specific, measurable failure point in clinical practice. For researchers working at the intersection of human factors and machine learning, this project offers a concrete model: define the error rate, isolate contributing factors through literature review, then design controlled experiments that measure how system design choices change expert behavior. The work also highlights a growing need for AI for Science & Research that accounts for human cognitive limitations rather than treating AI output as an endpoint. The question is not simply whether an algorithm detects nodules, but how the information it provides reshapes a trained specialist's visual search strategy - and whether that reshaping helps or harms the final diagnosis.
Your membership also unlocks:
