Development of Periapical Index Score Classification System in Periapical Radiographs Using Deep Learning

Area of Science:

• Dentistry
• Radiology
• Artificial Intelligence

Background:

• The Periapical Index (PAI) scoring system is standard for evaluating apical periodontitis (AP) on radiographs.
• Manual PAI scoring is time-consuming and requires dental expertise.
• Deep learning models show promise in automating PAI scoring but struggle with early AP stages.

Purpose of the Study:

• To develop and compare binary classification methods for PAI scoring using deep learning.
• To evaluate the efficacy of normality (PAI 1 vs. others) and health-disease (PAI 1-2 vs. others) classification approaches.
• To determine the optimal deep learning strategy for accurate AP assessment.

Main Methods:

• Utilized GoogLeNet, AlexNet, and ResNet convolutional neural networks (CNNs).
• Trained models on 2266 periapical root areas (PRAs) from 520 periapical radiographs (PRs).
• Compared normality classification (PAI 1=Normal, PAI 2-5=Abnormal) against health-disease classification (PAI 1-2=Healthy, PAI 3-5=Diseased).

Main Results:

• The normality classification method achieved a maximum accuracy of 75.00%.
• The health-disease classification method demonstrated superior performance with a maximum accuracy of 83.33%.
• CNN models showed improved classification accuracy when PAI scores 1 and 2 were grouped.

Conclusions:

• Grouping PAI scores 1 and 2 as 'Healthy' enhances deep learning model performance.
• The health-disease classification approach is more effective for automated PAI scoring.
• Findings support the clinical utility of the health-disease PAI scoring method for AP evaluation.