Machine classification of dental images with visual search

Area of Science:

• Dental radiology
• Machine learning
• Computer-assisted diagnosis

Background:

• Accurate diagnosis of periapical disease is crucial in dentistry.
• Human interpretation of dental radiographs can be subjective and prone to error.
• Developing objective, automated diagnostic tools is an ongoing area of research.

Purpose of the Study:

• To evaluate a computer-based classification system for periapical disease.
• To assess the performance of machine learning using observer gaze data to define image features.
• To compare the accuracy of machine classification with human observer accuracy.

Main Methods:

• Thirty-two dental radiographs with varying degrees of periapical disease were classified by an expert.
• Six observers independently classified images while their eye gaze was recorded.
• A machine classifier was developed using image space defined by visual gaze, random, and constrained random selection techniques.
• Kappa analyses compared classification accuracies.

Main Results:

• The machine classifier achieved 84% accuracy, significantly outperforming human observers (57% accuracy).
• Gaze-selected feature space yielded superior machine classification accuracy (kappa = 0.78) compared to random selection methods (kappa = 0.69, 0.68).
• Observer accuracy (kappa = 0.44) was substantially lower than both machine classification methods.

Conclusions:

• Machine classification of dental images, utilizing gaze-selected image space, enhances diagnostic accuracy.
• This computer-based system offers a promising tool to augment the diagnostic capabilities of dental professionals.
• Observer gaze data can effectively define relevant image features for machine learning in dental radiology.