Deep Learning-Based Semantic Segmentation and Classification of Otoscopic Images for Otitis Media Diagnosis and...

Area of Science:

• Medical imaging and artificial intelligence
• Otolaryngology and diagnostic technology

Background:

• Otitis media (OM), encompassing acute otitis media (AOM) and chronic otitis media (COM), is a prevalent condition requiring accurate diagnosis.
• Subjectivity in otoscopic interpretation necessitates objective diagnostic support, which artificial intelligence (AI) can provide through automated image analysis.

Purpose of the Study:

• To develop and evaluate an AI-based diagnostic framework for automated otitis media detection from clinical otoscopic images.
• To enhance diagnostic accuracy and consistency in identifying normal ears, AOM, and COM.

Main Methods:

• A three-step AI framework was developed: semi-supervised learning for tympanic membrane segmentation, region-based feature extraction, and disease classification.
• Utilized 607 clinical otoscopic images (normal, AOM, COM) for training (485) and testing (122) with U-Net and other CNN architectures for segmentation.
• Extracted color and texture features from segmented regions to train a classifier for disease state differentiation.

Main Results:

• The U-Net model achieved 96.76% pixel accuracy and 71.68% Dice coefficient for semantic segmentation of tympanic membrane structures.
• The AI framework demonstrated high diagnostic accuracies: 100% for normal ears, 100% for AOM, and 91.3% for COM on the test set.
• Overall diagnostic accuracy for the AI framework reached 96.72%.

Conclusions:

• A semi-supervised, segmentation-driven AI pipeline effectively integrates feature extraction and classification for high-accuracy otitis media diagnosis.
• The proposed AI framework offers a clinically interpretable, automated solution to improve diagnostic consistency and support clinical decision-making.
• This technology has the potential to facilitate scalable otoscopic assessment in various healthcare settings for disease prevention.