Optimization of MobileNet Architecture with Ghost Module for Dental Enamel Caries Classification

Abstract

Dental and oral diseases, particularly dental caries, represent a global health issue that requires early identification at the enamel stage to prevent further demineralization and damage. The utilization of artificial intelligence technology through Convolutional Neural Networks (CNNs) has been widely applied for medical image analysis. However, complex conventional models often incur high computational loads. Therefore, this study aims to implement and evaluate MobileNet variants (V1, V2, V3, and V4) optimized using the Ghost Module to classify dental caries images. The integration of the Ghost Module aims to mitigate feature redundancy and enhance feature representation without compromising image extraction quality. A dataset of 2,000 clinical dental images from the public "Caries-Spectra," categorized as advanced enamel caries, early-stage enamel caries, and no enamel caries, was curated and expanded to 12,000 images using preprocessing and augmentation. The dataset splitting was executed with a final learning proportion of 80% training data, 10% validation data, and 10% testing data. The image preprocessing utilized histogram equalization, CLAHE, and adaptive thresholding methods. The overall hybrid architecture was then evaluated based on performance metrics (such as accuracy, precision, recall, and F1 score). The experimental results demonstrate that the integration of the Ghost Module consistently enhances the classification accuracy across all MobileNet variants. The proposed Hybrid MobileNetV1 with Ghost Module achieved the highest performance, securing an overall accuracy of 96.83%, with well-balanced precision and recall across all enamel caries stages