Accurate detection of impacted teeth in panoramic radiographs plays a vital role in the diagnosis and treatment planning process within the field of dentistry. Traditional image segmentation methods often struggle with accurately identifying dental structures due to the complex and variable nature of human teeth. This study presents a comprehensive comparative analysis of advanced pixel-based segmentation models aimed at enhancing the accuracy of impacted teeth detection. Specifically, the study evaluates three models: UNet2, UNet3, and a hybrid model combining UNet with a Transformer architecture. The models are tested on a dataset of children's dental panoramic radiographs, sourced from the Kaggle dataset "Children's Dental Panoramic Radiographs," which contains high-resolution images with varying degrees of tooth impaction. The evaluation focuses on model performance across several key metrics, including segmentation accuracy, precision, recall, and F1-score. Results indicate that the hybrid UNet-Transformer model outperforms traditional models in terms of segmentation accuracy, demonstrating the potential of combining convolutional neural networks (CNNs) and attention mechanisms for complex image segmentation tasks. This research aims to provide a robust solution for improving automated dental diagnosis and support clinicians in making informed treatment decisions.

Keywords: Panoramic radiographs, impacted teeth detection, image segmentation, UNet, UNet3, hybrid model, Transformer, dental radiographs, machine learning, convolutional neural networks (CNN), attention mechanisms.