End-to-end CNN-based detection of permanent first molars and prediction of root development stages from panoramic radiographs

Abstract

The aim of this study was to develop a convolutional neural network (CNN)-based end-to-end learning architecture to predict the root development stages of permanent first molar teeth using panoramic radiographs. A dataset of 1629 first molar images was labeled according to the Cvek classification and organized into five subsets (DB-1 to DB-5) based on root development stages and apical foramen status. Teeth patches were cropped using the YOLO approach, and stage prediction was performed with VGG-19, InceptionV3, and EfficientNet-B3 models optimized with the Adamax optimizer at a learning rate of 10-3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10 {-3}$$\end{document}. The proposed method achieved high precision (98.4%) and recall (97.6%) in detecting first molar teeth. Classification performance reached average accuracies of 64.21% for DB-1, 62.66% for DB-2, and 69.64% for DB-3. For apical foramina classification, an accuracy of 84.57% was obtained in DB-4, which further improved to 94.89% in DB-5. These findings highlight the potential of CNN-based approaches in dental diagnostics, providing clinicians with an effective tool for assessing root development and supporting treatment planning.