Transformer-based classification with enhanced causal explainability from otoscopic images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Otolaryngology

Background:

• Otitis media is a leading cause of hearing loss, especially in children.
• Diagnosis is challenging due to nonspecific symptoms and subjective assessments.
• Current diagnostic methods lack transparency and reliability.

Purpose of the Study:

• To develop and evaluate transformer-based models for classifying tympanic membrane conditions from otoscopic images.
• To enhance diagnostic transparency and reliability in clinical settings using explainable AI.
• To assess the performance of Vision Transformer (ViT) and Data-efficient Image Transformer (DeiT) models.

Main Methods:

• Trained ViT and DeiT models on 454 pediatric and adult otoscopic images for multi-class classification (normal, effusion, tube).
• Employed Gradient-weighted Class Activation Map (Grad-CAM), Layer-wise Relevance Propagation (LRP), and Attention Rollout (AR) for explainability.
• Utilized a hybrid fusion strategy with Canonical Correlation Analysis and evaluated using insertion/deletion causal metrics.

Main Results:

• ViT model achieved 97.78% accuracy (AUC: 0.998), outperforming DeiT (93.33% accuracy, AUC: 0.994).
• ViT model achieved a 97.30% F1-score for the effusion class.
• Hybrid LRP and AR methods provided superior explainability, accurately highlighting critical image features.

Conclusions:

• Transformer-based models integrated with hybrid explainability methods significantly improve diagnostic transparency for otitis media.
• These advancements build clinician trust and establish a foundation for reliable clinical decision support systems.
• The study demonstrates the potential of AI in enhancing the accuracy and reliability of otitis media diagnosis.