Feasibility of the Machine Learning Network to Diagnose Tympanic Membrane Lesions without Coding Experience

Area of Science:

• Otolaryngology
• Medical Imaging
• Machine Learning

Background:

• Teachable Machine is a user-friendly, no-code machine learning platform.
• Accurate diagnosis of tympanic membrane lesions is crucial for effective treatment.

Purpose of the Study:

• To evaluate the performance of Teachable Machine for diagnosing tympanic membrane lesions.
• To assess the accuracy of a machine learning model trained without coding expertise.

Main Methods:

• Trained a classification network using 3024 tympanic membrane images labeled as normal, otitis media with effusion (OME), chronic otitis media (COM), or cholesteatoma.
• Evaluated performance across three classification levels: normal vs. abnormal (Level I), normal/OME/COM+cholesteatoma (Level II), and all four pathologies (Level III).
• Tested the model on 80 representative images.

Main Results:

• Mean accuracy for Level I (normal vs. abnormal) was 90.1%.
• Mean accuracy for Level II was 89.0%, and for Level III (all four pathologies) was 86.2%.
• Overall accuracy on test images was 78.75%, with specific hit rates for normal (95.0%), OME (70.0%), COM (90.0%), and cholesteatoma (60.0%).

Conclusions:

• Teachable Machine can effectively generate diagnostic networks for tympanic membrane classification.
• The no-code approach shows promise for developing AI tools in otology.
• Further validation may be needed for complex classifications like cholesteatoma.