▲ Professor Joong Ho Ahn from the Department of Otorhinolaryngology–Head and Neck Surgery at Asan Medical Center examines the ear of a patient with otitis media using an endoscope.
A recent study has found that using artificial intelligence (AI) to analyze otoscopic images can improve physicians’ diagnostic accuracy for ear diseases, including otitis media. The benefits of AI-assisted diagnosis were particularly pronounced among less experienced residents.
A research team led by Professors Joong Ho Ahn from the Department of Otorhinolaryngology-Head and Neck Surgery and Jihoon Kweon from the Department of Convergence Medicine at Asan Medical Center developed an AI model trained on 7,335 otoscopic images and evaluated its effectiveness in supporting diagnoses by 13 physicians. The results showed that junior residents’ accuracy in distinguishing chronic otitis media from otitis media with effusion and identifying cases involving neither condition increased by 32.2 percentage points, from 54.4% without AI assistance to 86.6% with it.
The diagnostic accuracy gap between ear disease specialists and junior residents also narrowed significantly, from 24.3 percentage points to just 3.1 percentage points with AI assistance. These findings suggest that AI can help reduce differences in diagnostic accuracy associated with physicians’ levels of clinical experience.
Middle ear diseases are conditions that affect the space behind the eardrum. Common examples include chronic otitis media and otitis media with effusion, in which fluid accumulates behind the eardrum. These conditions are diagnosed by examining the eardrum with an otoscope. However, their clinical features can vary, and some findings are difficult to distinguish, making diagnostic accuracy dependent on physicians’ experience and expertise.
To address these diagnostic challenges, the research team developed a deep learning-based AI model. Of the 7,435 otoscopic images taken at Asan Medical Center between 2018 and 2020, 7,335 were used to develop the AI model, while the remaining 100 were used to evaluate physicians’ diagnostic accuracy.
The AI model was designed to analyze otoscopic images and distinguish between chronic otitis media, otitis media with effusion, and cases involving neither condition. It was also trained to simultaneously identify additional conditions and eardrum features, including attic retraction, in which the upper part of the eardrum is pulled inward; myringitis, which is inflammation of the eardrum; otomycosis, a fungal infection of the ear; and the presence or absence of a tympanostomy tube. The AI presents its predicted diagnoses, which physicians then use as a reference when making their final clinical decisions.
In cross-validation using 7,335 otoscopic images, the AI model achieved 94.8% diagnostic accuracy across key categories, including the differentiation of chronic otitis media and otitis media with effusion. It also achieved 95% accuracy on a separate set of 100 evaluation images.
The research team then assessed the diagnostic support provided by AI among 13 physicians: three professors specializing in ear diseases, five senior residents, and five junior residents. First, the physicians examined 100 otoscopic images and made diagnoses without AI assistance. Three months later, they reviewed the same images and made diagnoses again, this time using the AI-generated results as a reference. The results showed statistically significant improvements in diagnostic accuracy across all physician groups for key categories, including the differentiation of chronic otitis media and otitis media with effusion.
Junior residents showed the greatest improvement in diagnostic accuracy, with their scores increasing by 32.2 percentage points, from 54.4% to 86.6%. Senior residents also showed a substantial improvement, from 65% to 93.8%, an increase of 28.8 percentage points. The professor group’s accuracy rose by 11 percentage points, from 78.7% to 89.7%. All three groups demonstrated improved diagnostic accuracy with AI assistance.
Improvements were also observed in specific diagnostic categories. Junior residents’ accuracy in diagnosing myringitis increased from 75.4% to 88.4%, while their accuracy in identifying attic retraction improved from 94.2% to 98.4%.
When all 13 physicians independently evaluated the 100 otoscopic images, 384 diagnoses (29.5%) that were initially incorrect were corrected after consulting the AI results. However, 46 diagnoses (3.5%) that were initially correct became incorrect after AI assistance.
“This study demonstrates that AI can help improve physicians’ diagnostic accuracy, particularly among those with less clinical experience, when used as a diagnostic support tool. With further validation involving diverse patient populations and clinical settings, we expect AI to be useful in diagnosing middle ear diseases and training medical professionals,” said Professor Joong Ho Ahn from the Department of Otorhinolaryngology-Head and Neck Surgery at Asan Medical Center.
The findings were recently published in JMIR Medical Informatics, an international journal in the field of medical informatics (Impact Factor: 5.0).