Multi-modal deep learning for joint prediction of otitis media and diagnostic difficulty

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Otolaryngology

Background:

• Otitis media is a common ear infection requiring accurate diagnosis.
• Current diagnostic methods may have limitations in accuracy and efficiency.
• Combining imaging and physiological measurements can potentially improve diagnostic performance.

Purpose of the Study:

• To develop and evaluate a novel diagnostic model for automatic otitis media detection.
• To investigate the efficacy of combining otoscopy images and wideband tympanometry data.
• To compare a deep metric learning approach with a standard multi-task network for joint prediction of otitis media and diagnostic difficulty.

Main Methods:

• A neural network-based model utilizing deep metric learning was developed.
• The model processed combined inputs of otoscopy images and wideband tympanometry measurements.
• Performance was evaluated for classifying otitis media (acute otitis media and otitis media with effusion) and estimating diagnostic difficulty.

Main Results:

• The multi-modal approach significantly improved performance compared to single-modality models.
• The deep metric learning model achieved 86.5% accuracy for otitis media classification.
• A Kendall rank correlation coefficient of 0.45 was obtained for difficulty estimation, correctly ranking 72.6% of cases.

Conclusions:

• A multi-modal diagnostic tool integrating otoscopy images and wideband tympanometry is effective for otitis media diagnosis.
• Deep metric learning enhances the performance of diagnostic models for otitis media.
• This approach offers a promising advancement in the automated detection and assessment of otitis media.