Generalizability of deep learning models for dental image analysis.

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Imaging Analysis
• Deep Learning for Diagnostic Support

Background:

• Deep learning models require robust generalizability for reliable clinical application.
• Detecting apical lesions on panoramic radiographs is crucial for dental diagnostics.
• Variations in imaging protocols and patient populations can hinder model generalizability.

Purpose of the Study:

• To assess the generalizability of deep learning models for apical lesion detection across different imaging centers.
• To investigate methods for improving the generalizability of these models.
• To explore the impact of image characteristics and dental status on model performance.

Main Methods:

• Trained U-Net deep learning models on panoramic radiographs from a German center (Charité).
• Evaluated model performance on test sets from both the German (Charité) and an Indian center (KGMU).
• Investigated cross-center training by incrementally including data from the Indian center and analyzed the influence of dental status.

Main Results:

• Models trained solely on German data performed significantly worse on Indian data (F1-score 32.7% vs. 54.1%).
• Aligning image characteristics did not enhance generalizability.
• Gradually incorporating Indian data into training improved performance on Indian images (F1-score 46.1%) with a slight decrease on German images.
• Model performance was notably lower for images lacking root-canal fillings or restorations.

Conclusions:

• Deep learning models for apical lesion detection exhibit limited generalizability across different centers.
• Cross-center training is an effective strategy to improve model generalizability.
• Patient dental status is a significant factor influencing model performance, more so than image characteristics.