Benchmarking Segmentation Models for Identifying Impacted Teeth in Panoramic Radiographs

The paper provides a new comparative performance of the pixel-based segmentation models to accurately identify the affected teeth in panoramic radiographical images. Using the analysis of the Dental Panoramic Radiographs dataset of Kaggle that provides the full set of images that will be specific to this task the work examines the work of five advanced segmentation algorithms: U-Net, DeepLabv3+, U-Net2, U-Net++, and SwinUNet. Intersection over Union (IoU) is the most important evaluation parameter and it is an effective measure of a segmentation accuracy. It is important to note that U-Net++ performs better than the rest with a score of 0.8838, and the U-Net2 is close behind with a score of 0.8734. On the contrary, such models as DeepLabv3+ and SwinUNet have relatively poor segmentation results. This research presents a very efficient method whereby the latest and advanced deep learning structures are used to recognize affected areas of the teeth using radiographs. The suggested methodology will provide a tremendous improvement in the precision and effectiveness of the dental practitioners in identifying the teeth affected and hence improve the automation of dental diagnosis. This project makes a significant contribution to the medical image field especially in the design of computerized systems to process panorama radiographies during dental practice.