Enhancing deep learning-based dental caries detection through feature pyramid network-based image segmentation.

Context: Dental caries is a multifactorial disease that may lead to pain, infection, and tooth loss if left untreated, and it requires early detection to prevent complications and to reduce costs. Advancement of artificial intelligence (AI) has enabled automated caries detection using intraoral images. However, many deep learning models rely on raw images, such as gums, tongue, and lighting variations, resulting in reduced diagnostic accuracy. Tooth region segmentation is therefore essential to focus on relevant features and to produce more reliable predictions. Aims: This study aims to improve the performance of classification models by employing a feature pyramid network (FPN)-based tooth segmentation method before classification. Settings and Design: The FPN technique was utilized for the teeth image segmentation model, which is integrated with feature extraction backbones. Materials and Methods: Three deep learning architectures, namely DenseNet169, InceptionV3, and MobileNetV2, were also employed for caries classification, and were trained with and without FPN-based segmentation. Segmentation was evaluated using IoU and Dice Similarity Coefficient for performance. Statistical Analysis Used: A comparative analysis was conducted on FPN-based segmentation by comparing models on segmented images and original images employing accuracy, precision, recall, and F1-score. Results: The findings show that employing FPN-based segmentation before the classification process consistently enhances the performance. DenseNet169 model outperformed InceptionV3 and MobileNetV2 models, yielding 89.28% of classification accuracy. Conclusions: The improvement in classification results also confirmed that employing segmentation techniques reinforces feature extraction, leading to more reliable classification, with strong potential for AI-based caries early screening using intraoral images.