A Machine Learning Approach to Screen for Otitis Media Using Digital Otoscope Images Labelled by an Expert Panel

Area of Science:

• Otolaryngology
• Medical Imaging
• Artificial Intelligence

Background:

• Otitis media, a common middle ear inflammation, can lead to severe complications if untreated.
• Accurate diagnosis of otitis media is challenging, highlighting the need for advanced screening tools.
• Machine learning (ML) offers potential for developing automated diagnostic systems for prevalent diseases.

Purpose of the Study:

• To evaluate the performance of a convolutional neural network (CNN) in screening for otitis media.
• To assess the CNN's accuracy in classifying digital otoscopic images.
• To compare ML-based classification with expert otologist diagnoses.

Main Methods:

• Five otologists diagnosed 347 digital otoscopic tympanic membrane images.
• Images with majority expert consensus (n=273) were classified into Normal, Pathological, or Wax categories.
• A CNN was trained and tested on these images, with various model approaches evaluated.

Main Results:

• The CNN achieved an overall accuracy exceeding 0.9 in most tested configurations.
• Sensitivity for detecting wax or pathology was consistently above 93%, with 100% specificity.
• Image augmentation and dataset normalization improved CNN performance in some instances.

Conclusions:

• Machine learning, specifically CNNs, demonstrates high accuracy in screening for otitis media from digital otoscopic images.
• This technology could serve as a valuable, objective screening tool in clinical settings.
• Further refinement of ML models may enhance diagnostic capabilities for middle ear conditions.