Deep Learning Applications in Cone-Beam Computed Tomography (CBCT) for Automatic Detection and Segmentation of Periapical Lesions: A Comprehensive Clinical and Radiographic Evaluation

Abstract

Proper diagnosis and segmentation of periapical lesions using Cone Beam Computed Tomography (CBCT) imaging play an important role in determining the best possible strategy for treatment of such pathologies. Even though CBCT provides a better three-dimensional visualization than regular two-dimensional X-ray images, multi-planar analysis still requires much time and depends on the observer's experience. Deep learning (DL) techniques, especially Convolutional Neural Networks (CNNs), provide an advanced tool that can help solve these issues. In the current research, the clinical performance, segmentation accuracy, and practical issues of modern DL methods are discussed, specifically the 3D DL models, such as 3D U-Net and ResNet, which were used on CBCT images from 2021 to 2026. High sensitivity (higher than 92%), high specificity, and small time needed to conduct the diagnostic procedure have been identified. However, some limitations, such as restoration artifacts, different sizes of voxels, and heterogeneity of datasets, still exist.