A recent study comparing the diagnostic abilities of artificial intelligence and human radiologists in detecting lung diseases from chest X-rays found that human radiologists outperformed AI. AI had a higher rate of false positives, incorrectly identifying diseases that weren’t there, when compared to human radiologists.
This study affirms the notion that although AI has the potential to provide valuable support in the field of medical diagnosis, its evolution is still in its nascent phase.
The study conducted by the Danish team at Herlev and Gentofte Hospital in Copenhagen compared the diagnostic abilities of 72 radiologists with four different AI tools using 2,040 adult X-rays from various Danish hospitals in 2020. About one-third (32.8%) of the X-rays revealed significant findings. The study focused on evaluating the real-world effectiveness of AI tools in diagnosing three common conditions in chest X-rays: airspace diseases, pneumothorax, and pleural effusion. This research addresses the growing popularity of AI tools in radiology amid a global shortage of radiologists.
AI tools achieved sensitivity rates ranging from 72% to 91% for airspace diseases, 63% to 90% for pneumothorax, and 62% to 95% for pleural effusion. Nonetheless, they exhibited higher false positive rates compared to human radiologists, particularly in identifying airspace diseases.
Herlev and Gentofte Hospital radiologist and lead researcher of the study Dr Louis Plesner said that chest radiography is a widely used diagnostic method, but it demands substantial expertise to interpret results accurately. He added that although AI tools are gaining acceptance in radiology, there is a pressing need to assess their performance in real clinical settings.
The researchers emphasize that, while AI can assist in the detection of typical chest X-ray findings, exclusively depending on it for diagnostic purposes is untimely. The team underscores that radiologists frequently surpass AI in various patient scenarios, indicating that AI’s present capability is not yet suitable for individual diagnoses. They contend that the future evolution of AI, which may integrate image analysis with patients’ medical histories and past imaging data, holds the promise of significantly enhanced diagnostic accuracy.