According to MIT researchers, As radiologists review medical images, they often rely on terms like “may,” “likely,” or “possibly” when describing potential pathologies in X-rays and scans. These terms are common because they convey the inherent uncertainty in medical imaging, but do they truly reflect the accuracy of the diagnosis? According to a groundbreaking study by MIT, the use of these phrases is more complex than previously thought.
The study, conducted by MIT researchers in collaboration with Harvard Medical School clinicians, explored how radiologists express their confidence in diagnoses using natural language. It turns out that certain words—such as “very likely”—can often signal overconfidence, while terms like “possibly” may reveal an underestimation of certainty. This research aims to create a more reliable framework for radiologists to improve their diagnostic language, with the ultimate goal of enhancing patient care.
Decoding Radiologists’ Confidence Levels
MIT’s research team developed a novel framework that quantifies the reliability of the confidence expressed in the words radiologists use. By analyzing clinical data and reports, they identified more precise ways to guide radiologists in their choice of words. For instance, the term “consistent with” could correspond to a higher likelihood of a condition, while “may represent” suggests a wider range of uncertainty.
The result? A calibration map that offers better guidance on how to use terms for various conditions. This new approach goes beyond simply attaching percentages to vague words like “likely.” Instead, it treats them as probability distributions that reflect the nuances in how radiologists interpret uncertainty. A term like “likely present” for a condition could improve the precision of a diagnosis, aligning better with the actual medical probabilities.
Enhancing Diagnostic Accuracy with AI
For the researchers, the big takeaway is that improving the accuracy of language used in diagnostics can ultimately lead to better patient outcomes. Peiqi Wang, a lead author and MIT graduate student, explained: “If practitioners can be more reliable in their reporting, patients will be the ultimate beneficiaries.”
By also applying this framework to AI systems, the team found that language models used in medical imaging often overstate their confidence, using terms like “certainly” that could discourage verification. For instance, AI might suggest a diagnosis with 100% certainty, but a closer look reveals uncertainty that was overlooked.
What’s Next for Medical Imaging?
The research doesn’t stop with improving the language of radiologists. MIT researchers plan to extend their work to abdominal CT scans and work on refining how radiologists use language to better express their confidence. There’s also ongoing exploration into whether radiologists can adapt to this more precise language by leveraging these suggestions in their daily practice.
Atul B. Shinagare, an associate professor at Harvard Medical School, emphasized the significance of this research for enhancing diagnostic accuracy. “This study provides a novel approach to improving diagnostic accuracy, offering feedback on how radiologists can better express their certainty, ultimately improving patient care.”
This study, supported by the MIT-IBM Watson AI Lab, marks a significant step in enhancing the reliability of medical imaging. By improving how radiologists communicate their confidence, the research promises to refine diagnostic processes, leading to faster, more accurate decisions and better outcomes for patients.
